2004/id/draft-holtman-http-negotiation-04.txt

HTTP Working Group                                     Koen Holtman, TUE
Internet-Draft                              Andrew Mutz, Hewlett-Packard
Expires: May 21, 1997                                  November 21, 1996


                 Transparent Content Negotiation in HTTP

                  draft-holtman-http-negotiation-04.txt


STATUS OF THIS MEMO

        This document is an Internet-Draft. Internet-Drafts are
        working documents of the Internet Engineering Task Force
        (IETF), its areas, and its working groups. Note that other
        groups may also distribute working documents as
        Internet-Drafts.

        Internet-Drafts are draft documents valid for a maximum of
        six months and may be updated, replaced, or obsoleted by
        other documents at any time. It is inappropriate to use
        Internet-Drafts as reference material or to cite them other
        than as "work in progress".

        To learn the current status of any Internet-Draft, please
        check the "1id-abstracts.txt" listing contained in the
        Internet-Drafts Shadow Directories on ftp.is.co.za
        (Africa), nic.nordu.net (Europe), munnari.oz.au (Pacific
        Rim), ds.internic.net (US East Coast), or ftp.isi.edu (US
        West Coast).

        Distribution of this document is unlimited.  Please send
        comments to the HTTP working group at
        <http-wg@cuckoo.hpl.hp.com>.  Discussions of the working
        group are archived at
        <URL:http://www.ics.uci.edu/pub/ietf/http/>.  General
        discussions about HTTP and the applications which use HTTP
        should take place on the <www-talk@w3.org> mailing list.

        A HTML version of this document, and versions with change
        bars, are available at
        <URL:http://gewis.win.tue.nl/~koen/conneg/>.


ABSTRACT

        HTTP allows one to put multiple versions of the same
        information under a single URL.  Transparent content
        negotiation is a mechanism, layered on top of HTTP, for
        automatically selecting the best version when the URL is
        accessed.  This enables the smooth deployment of new web data
        formats and markup tags.

        Design goals for transparent content negotiation include: low
        overhead on the request message size, downwards compatibility,
        extensibility, enabling the rapid introduction of new areas of
        negotiation, scalability, low cost for minimal support, end
        user control, and good cacheability.


TABLE OF CONTENTS

   Note: the section numbering is this document is sometimes irregular
   because it tries to preserve the numbering in the 02 version as
   much as possible.

       1  Introduction
        1.1 Background
        1.2 Note for readers
        1.3 Revision History

       2  Terminology
        2.1 Terms from HTTP/1.1
        2.2 New terms

       3  Notation

       4  Overview
        4.1 Content negotiation
        4.2 HTTP/1.0 style negotiation scheme
        4.3 Transparent content negotiation scheme
        4.4 Optimizing the negotiation process
        4.4B Downwards compatibility with non-negotiating user agents
        4.5 Retrieving a variant by hand
        4.6 Dimensions of negotiation
        4.7 Feature negotiation

       5  Variant descriptions
        5.1 Syntax
        5.2 URI
        5.3 Source-quality
        5.4 Type, language, length, and charset.
        5.5 Features
        5.6 Description
        5.7 Extension-attribute

       6  Feature negotiation
        6.1 Feature tags
        6.2 Accept-Features header
        6.3 Feature predicates
        6.4 Features attribute

       7  Feature negotiation examples
        7.1 Use of feature tags
        7.2 Use of numeric feature tags

       8  Feature tag registration
        8.1 Evolution of feature tags
        8.2 Core set of feature tags
        8.3 Feature tag design

       9  Content negotiation response codes and headers
        9.1 506 Variant Also Negotiates
        9.2 Accept-Features
        9.3 Content-Features
        9.4 Alternates
        9.6 Negotiate
        9.7 Variant-Vary

       9B Cache validators
        9B.1 Variant list validators
        9B.2 Structured entity tags
        9B.3 Assigning entity tags to variants

       10 Content negotiation responses
        10.1 List response
        10.2 Choice response
        10.2B Ad hoc response
        10.3 Reusing the Alternates header
        10.4 Extracting a normal response from a choice response
        10.5 Elaborate Vary headers
        10.5.1 Construction of an elaborate Vary header
        10.5.2 Caching of an elaborate Vary header
        10.6 Negotiation on content encoding

       11 The network negotiation algorithm
        11.1 Input
        11.2 Output
        11.2.1 Output for proxies
        11.2.2 Output for origin servers
        11.2.3 Output for user agents
        11.3 Computing the overall quality values
        11.4 Definite and speculative quality values
        11.5 Determining the result
        11.5.1 Result for proxies and origin servers
        11.5.2 Result for user agents
        11.6 Construction of short requests

       12 User agent support for transparent negotiation
        12.1 Handling of responses
        12.2 Presentation of a transparently negotiated resource

       13 Origin server support for transparent negotiation
        13.1 Requirements
        13.2 Negotiation on transactions other than GET and HEAD

       14 Proxy support for transparent negotiation

       15 Security and privacy considerations
        15.1 Accept headers revealing information of a private nature
        15.2 Spoofing of responses from variant resources

       16 Acknowledgments

       17 References

       18 Authors' addresses

       19 Appendices
        19.1 Adding an Expires header to ensure HTTP/1.0 compatibility
        19.2 Origin server implementation considerations
        19.2.1 Implementation with a CGI script
        19.2.2 Direct support by HTTP servers
        19.2.3 Web publishing tools
        19.3 Open issues in transparent content negotiation
        19.3B Server-side rendering engines
        19.3C Quality factors in the Accept-Features header
        19.4 Other negotiation specifications
        19.4.1 User-Agent Display Attributes
        19.4.2 HTTP/1.2 Extension Protocol (PEP)
        19.5 Related issues
        19.5.1 Current negotiation practice
        19.5.2 Bandwidth negotiation


1  Introduction

   HTTP allows one to put multiple versions of the same information
   under a single URI.  Each of these versions is called a `variant'.
   Transparent content negotiation is a mechanism for automatically
   and efficiently retrieving the best variant when a GET or HEAD
   request is made.  This enables the smooth deployment of new web
   data formats and markup tags.

   This specification defines transparent content negotiation as an
   extension on top of the HTTP/1.1 protocol [1].  However, use of
   this extension does not require use of HTTP/1.1: transparent
   content negotiation can also be done if some or all of the parties
   are HTTP/1.0 [3] systems.

   Transparent negotiation is called `transparent' because it makes
   all variants inside the origin server visible to outside parties.


1.1 Background

   The addition of content negotiation to the web infrastructure has
   been considered important since the early days of the web.  Among
   the expected benefits of a sufficiently powerful system for content
   negotiation are

     * smooth deployment of new data formats and markup tags will
       allow graceful evolution of the web

     * eliminating the need to choose between a `state of the art
       multimedia homepage' and one which can be viewed by all web
       users

     * enabling good service to a wider range of browsing
       platforms (from low-end PDA's to high-end VR setups)

     * eliminating error-prone and cache-unfriendly
       User-Agent based negotiation

     * enabling construction of sites without `click here for the X
       version' links

     * internationalization, and the ability to offer multi-lingual
       content without a bias towards one language.

   The development of content negotiation has been a slow process,
   perhaps because the expected benefits are mostly long-term
   benefits.  Also, all immediately obvious content negotiation
   solutions involve sending larger HTTP request messages.  Larger
   request messages are unacceptable to many people: a small overhead
   on the request size is often cited as the main technical
   requirement for any scheme which offers the above benefits.

   Therefore, a low overhead on the request message size is one of the
   main design goals of this specification. A large fraction of the
   text below is devoted to the definition of mechanisms for meeting
   this design goal.  The desire to provide for good cacheability in
   all cases accounts for another considerable part of the content of
   this document.

   Other design goals include downwards compatibility, extensibility,
   enabling the rapid introduction of new areas of negotiation,
   scalability, low cost of minimal support, and end user control.
   End user control, the option to manually select a variant if
   desired, is the one redeeming quality of the `click here for the X
   version' type of negotiation. Preserving this option is considered
   important.


1.2 Note for readers

   Some sections in this document discuss requirements for proxy
   caches only.  Implementers of origin servers and user agents (but
   not implementers of user agent caches) can skip these sections.
   The sections which can be skipped are:

       10.3 Reusing the Alternates header
       10.4 Extracting a normal response from a choice response
       10.5.2 Caching of an elaborate Vary header
       14 Proxy support for transparent negotiation

   The skipping of these sections is also recommended on first
   reading.


1.3 Revision History

   The change bars in this document reflect the changes with respect
   to the 03 version of this draft.  Major functional changes are:

    - The model for the caching of variant lists has been changed
      extensively. (The previous version had a problem with downwards
      compatibility.)

    - The definition of the Negotiate header has been changed to
      include a `trans' keyword.

    - The form `ftag=<N-M>', with N-M a range of numeric values, was
      added to the feature expressions and feature predicates.

   Major changes in the presentation are:

    - Feature tag registration is now covered in a separate draft

    - Some new explanatory material was added

    - Some existing text has been improved

   In the text version of this document with change bars, all changes
   are marked except changes in formatting and punctuation.

   In the HTML version of this document with changes marked, all
   changed words and symbols are typeset in bold text.  Deletions and
   changes in punctuation are not marked in the HTML version.


2  Terminology

2.1 Terms from HTTP/1.1

   This specification mostly uses the terminology of the HTTP/1.1
   specification [1].  The definitions below were reproduced from [1].

   request
     An HTTP request message.

   response
     An HTTP response message.

   resource
     A network data object or service that can be identified by a URI.
     Resources may be available in multiple representations
     (e.g. multiple languages, data formats, size, resolutions) or
     vary in other ways.

   content negotiation
     The mechanism for selecting the appropriate representation when
     servicing a request.

   variant
     A resource may have one, or more than one, representation(s)
     associated with it at any given instant.  Each of these
     representations is termed a `variant.'  Use of the term `variant'
     does not necessarily imply that the resource is subject to
     content negotiation.

   client
     A program that establishes connections for the purpose of sending
     requests.

   user agent
     The client which initiates a request.  These are often browsers,
     editors, spiders (web-traversing robots), or other end user
     tools.

   server
     An application program that accepts connections in order to
     service requests by sending back responses.  Any given program may
     be capable of being both a client and a server; our use of these
     terms refers only to the role being performed by the program for
     a particular connection, rather than to the program's
     capabilities in general.  Likewise, any server may act as an
     origin server, proxy, gateway, or tunnel, switching behavior
     based on the nature of each request.

   origin server
     The server on which a given resource resides or is to be created.

   proxy
     An intermediary program which acts as both a server and a client
     for the purpose of making requests on behalf of other
     clients.  Requests are serviced internally or by passing them on,
     with possible translation, to other servers.  A proxy must
     implement both the client and server requirements of this
     specification.

   age
     The age of a response is the time since it was sent by, or
     successfully validated with, the origin server.

   fresh
     A response is fresh if its age has not yet exceeded its freshness
     lifetime.

2.2 New terms

   transparently negotiable resource
     A resource, identified by a single URI, which has multiple
     representations (variants) associated with it.  When servicing a
     request on its URI, it allows selection of the best
     representation using the transparent content negotiation
     mechanism.  A transparently negotiable resource always has a
     variant list bound to it, which can be represented as an
     Alternates header.

   variant list
     A list containing variant descriptions, which can be bound to a
     transparently negotiable resource.

   variant description
     A machine-readable description of a variant resource, usually
     found in a variant list.  A variant description contains the
     variant resource URI and various attributes which describe
     properties of the variant.  Variant descriptions are defined in
     Section 5.

   variant resource
     A resource from which a variant of a negotiable resource can be
     retrieved with a simple GET request.  A variant resource usually
     has only one representation (variant) associated with it.

   list response
     A list response contains the variant list of the negotiable
     resource, but no variant data.  It is generated when the server
     does not (perhaps cannot) choose a particular best variant for the
     request.  List responses are defined in Section 10.1.

   choice response
     A choice response contains both the variant list of the
     negotiable resource and a representation of the best variant for
     the request.  Choice responses are defined in Section 10.2.

   ad hoc response
     An ad hoc response contains the variant list of the negotiable
     resource, and any other data the origin server wants to send.  It
     can be generated as a response to a non-negotiating user agent if
     the server does not (perhaps cannot) choose any particular
     variant. Ad hoc responses are defined in Section 10.2B.

   normal response
     A HTTP response which is not a list response, choice response, or
     ad hoc response.  Normal responses are never generated by
     transparently negotiable resources, and are always generated by
     variant resources.

   Accept headers
     The request headers: Accept, Accept-Charset, Accept-Language, and
     Accept-Features.

   network negotiation algorithm
     A standardized algorithm by which a party in the negotiation
     process can choose a best variant on behalf of another party.
     The algorithm computes whether the Accept headers in the request
     contain sufficient information to allow a choice, and if so,
     which variant must be chosen.  The network negotiation algorithm
     is defined in Section 11.

   neighbor
     Two resources are called neighbors if the absolute URI of the
     first resource up to its last slash equals the absolute URI of
     the second resource up to its last slash.  The neighboring
     relation is important because of security considerations; see
     Section 15.2.

   [##Historical note: In the context of old versions of the HTTP/1.1
   specification like [2], list responses can be called `pre-emptive
   negotiation responses', and choice responses can be called
   `reactive negotiation responses'.##]

   [##Note: In the context of the HTTP/1.1 specification [1], both
   list and choice responses are the result of server-driven
   negotiation, though a list response is a server-driven negotiation
   response meant to initiate agent-driven negotiation.##]


3  Notation

   The notation [## ... ##] in this document encloses an editorial
   comment.  Such a comment will be either removed or replaced by real
   text in the final version of this document.

   The version of BNF used in this document is taken from [1], and
   many of the nonterminals used are defined in [1].

   One new BNF construct is added:

      1%rule

   stands for one or more instances of "rule", separated by
   whitespace:

      1%rule =  rule *( 1*LWS rule )


4  Overview

   This section gives an overview of transparent content negotiation.
   It starts with a more general discussion of negotiation as provided
   by HTTP.


4.1 Content negotiation

   HTTP/1.1 allows one to put multiple versions of the same
   information under a single resource URI.  Each of these versions is
   called a `variant'. For example, a resource http://x.org/paper
   could bind to three different variants of a paper:

         1. HTML, English
         2. HTML, French
         3. Postscript, English

   Content negotiation is the process by which the best variant is
   selected if the resource is accessed.  The selection is done by
   matching the properties of the available variants to the
   capabilities of the user agent and the preferences of the user.

   It has always been possible under HTTP to have multiple
   representations available for one resource, and to return the most
   appropriate representation for each subsequent request.
   HTTP/1.1 is the first version of HTTP which has provisions for
   doing this in a cache-friendly way.  These provisions include the
   Vary response header, entity tags, and the If-None-Match request
   header.


4.2 HTTP/1.0 style negotiation scheme

   The HTTP/1.0 protocol elements allow for a negotiation scheme as
   follows:

      Server _____ proxy _____ proxy _____ user
      x.org        cache       cache       agent

        < ----------------------------------
        |      GET http://x.org/paper
        |          Accept headers
      choose
        |
         ---------------------------------- >
                    Best variant

   When the resource is accessed, the user agent sends (along with its
   request) various Accept headers which express the user agent
   capabilities and the user preferences.  Then the origin server uses
   these Accept headers to choose the best variant, which is returned
   in the response.

   The biggest problem with this scheme is that it does not scale
   well.  For all but the most minimal user agents, Accept headers
   expressing all capabilities and preferences would be very large,
   and sending them in every request would be hugely inefficient, in
   particular because only a small fraction of the resources on the
   web have multiple variants.


4.3 Transparent content negotiation scheme

   The transparent negotiation scheme eliminates the need to send huge
   Accept headers, and nevertheless allows for a selection process
   that always yields either the best variant, or an error message
   indicating that user agent is not capable of displaying any of the
   available variants.

   Under the transparent content negotiation scheme, the server sends
   a list with the available variants and their properties to the user
   agent.  An example of a list with three variants is

     {"paper.html.en" 0.9 {type text/html} {language en}},
     {"paper.html.fr" 0.7 {type text/html} {language fr}},
     {"paper.ps.en"   1.0 {type application/postscript} {language en}}

   The syntax and semantics of the variant descriptions in this list
   are covered in depth in Section 5.  When the list is received, the
   user agent can choose the best variant and retrieve it.
   Graphically, the communication can be represented as follows:

      Server _____ proxy _____ proxy _____ user
      x.org        cache       cache       agent

        < ----------------------------------
        |      GET http://x.org/paper
        |
        ----------------------------------- >         [list response]
                  return of list            |
                                         choose
                                            |
        < ----------------------------------
        |  GET http://x.org/paper.html.en
        |
         ---------------------------------- >         [normal response]
                return of html.en

   The first response returning the list of variants is called a `list
   response'.  The second response is a normal HTTP response: it does
   not contain special content negotiation related information.  Only
   the user agent needs to know that the second request actually
   retrieves a variant, for the other parties in the communication,
   the second transaction is indistinguishable from a normal HTTP
   transaction.

   With this scheme, information about capabilities and user
   preferences is only used by the user agent itself.  Therefore,
   sending such information in large Accept headers is unnecessary.
   Accept headers do have a limited use in transparent content
   negotiation however; the sending of small Accept headers can often
   speed up the negotiation process. This is covered in Section 4.4.

   List responses are covered in depth in Section 10.1.  As an
   example, the list response in the above picture could be:

     HTTP/1.1 300 Multiple Choices
     Date: Tue, 11 Jun 1996 20:02:21 GMT
     Alternates: {"paper.html.en" 0.9 {type text/html} {language en}},
                 {"paper.html.fr" 0.7 {type text/html} {language fr}},
                 {"paper.ps.en"   1.0 {type application/postscript}
                     {language en}}
     Vary: negotiate, accept, accept-language
     ETag: "blah;1234"
     Cache-control: max-age=86400
     Content-Type: text/html
     Content-Length: 227

     <h2>Multiple Choices:</h2>
     <ul>
     <li><a href=paper.html.en>HTML, English version</a>
     <li><a href=paper.html.fr>HTML, French version</a>
     <li><a href=paper.ps.en>Postscript, English version</a>
     </ul>

   The Alternates header in the response contains the variant list.
   The Vary header is included to ensure correct caching by HTTP/1.1
   caches not capable of transparent content negotiation (see Section
   10.5).  The ETag header allows the response to be revalidated by
   caches, the Cache-Control header controls this revalidation.  The
   HTML entity included in the response allows the user to select the
   best variant by hand if desired.


4.4 Optimizing the negotiation process

   The basic transparent negotiation scheme involves two HTTP
   transactions: one to retrieve the list, and a second one to retrieve
   the chosen variant.  There are however several ways to `cut corners'
   in the data flow path of the basic scheme.

   First, caching proxies can cache both variant lists and complete
   responses.  Such caching can reduce the communication overhead, as
   shown in the following example:

      Server _____ proxy _____ proxy __________ user
      x.org        cache       cache            agent

                                 < --------------
                                 |  GET ../paper
                                 |
                               has the list
                               in cache
                                 |
                                  -------------  >  [list response]
                                           list  |
                                                 |
                                              choose
                                                 |
                     < --------------------------
                     |   GET ../paper.html.en
                     |
                  has the variant
                  in cache
                     |
                      -------------------------- >  [normal response]
                         return of html.en

   Second, the user agent can send small Accept headers which could
   allow a server to determine the choice the user agent would make on
   receiving the list.  If the Accept headers contain enough
   information, the origin server can send back the variant directly:

      Server _____ proxy _____ proxy _____ user
      x.org        cache       cache       agent

        < ----------------------------------
        |      GET http://x.org/paper
        |       small Accept headers
        |
      able to choose on
      behalf of user agent
        |
         ---------------------------------- >    [choice response]
              return of html.en and list

   This choosing based on small accept headers can be done with the
   network negotiation algorithm (Section 11).  This algorithm takes
   the variant list and the Accept headers as input.  It computes
   whether the Accept headers contain sufficient information to choose
   on behalf of the user agent, and if so, which variant must be
   chosen.

   The response in the above diagram is called a choice response. It
   transmits both the chosen variant and the list of all variants
   bound to the negotiable resource.  Choice responses are covered in
   depth in Section 10.2.  For example, the choice response in the
   above picture could be:

     HTTP/1.1 200 OK
     Date: Tue, 11 Jun 1996 20:05:31 GMT
     Content-Type: text/html
     Last-Modified: Mon, 10 Jun 1996 10:01:14 GMT
     Content-Length: 5327
     Cache-control: max-age=604800
     Content-Location: paper.html.en
     Alternates: {"paper.html.en" 0.9 {type text/html} {language en}},
                 {"paper.html.fr" 0.7 {type text/html} {language fr}},
                 {"paper.ps.en"   1.0 {type application/postscript}
                     {language en}}
     Etag: "gonkyyyy;1234"
     Vary: negotiate, accept, accept-language
     Expires: Thu, 01 Jan 1980 00:00:00 GMT

     <title>A paper about ....

   Finally, the above two kinds of optimization can be combined; a
   caching proxy which has the list will sometimes be able to choose on
   behalf of the user agent.  This could lead to the following
   communication pattern:

      Server _____ proxy _____ proxy __________ user
      x.org        cache       cache            agent

                                 < ---------------
                                 |  GET ../paper
                                 |  small Accept
                                 |
                              able to choose
                                on behalf
                                 |
                     < ----------
                     |  GET ../paper.html.en
                     |
                      ---------- >   [normal response]
                        html.en  |
                                  ---------------- >  [choice response]
                                   html.en and list

   In this case the proxy cache nearest to the user agent has
   a cached variant list, and chooses on behalf of the user agent.

   Note that this cutting of corners does not only save bandwidth. It
   also eliminates delays due to packet round trip times, and reduces
   the load on the origin server.


4.4B Downwards compatibility with non-negotiating user agents

   To handle requests from user agents not capable of transparent
   content negotiation, transparent content negotiation basically
   reverts to a HTTP/1.0 style negotiation scheme.  One optimization
   remaining is that proxy caches may sometimes be able to perform
   HTTP/1.0 style negotiation on behalf of the origin server.


4.5 Retrieving a variant by hand

   If a transparently negotiated resource is accessed, the user agent
   will always at some point receive the list of available variants.
   The user agent can use this list to make available a menu (in HTML or
   not) of all variants and their characteristics to the user.
   Such a menu can allow the user to randomly browse other variants,
   and can make it possible to manually correct any sub-optimal
   choice made by the automatic negotiation process.


4.6 Dimensions of negotiation

   Transparent content negotiation defines four dimensions of
   negotiation:

     1. Media type (MIME type)
     2. Charset
     3. Language
     4. Features

   The first three dimensions have traditionally been present in HTTP.
   The fourth dimension is added by this specification.  Additional
   dimensions, beyond the four mentioned above, could be added by
   future specifications.

   Negotiation on the content encoding of a response (gzipped,
   compressed, etc.) is left outside of the realm of transparent
   negotiation.   See Section 10.6 for more information.

     [##Note: In earlier versions of this specification, content
     encoding was a regular dimension in transparent content
     negotiation.  It was removed mainly because negotiation on
     content encoding can be done on a hop-by-hop basis under HTTP/1.1
     [1].##]


4.7 Feature negotiation

   Feature negotiation intends to provide for all areas of negotiation
   not covered by the type, charset, and language dimensions.
   Examples are negotiation on

       * HTML extensions
       * Extensions of other media types
       * Color capabilities of the user agent
       * Screen size
       * Output medium (screen, paper, ...)
       * Preference for speed vs. preference for graphical detail

   The feature negotiation framework is the principal means by which
   transparent negotiation offers extensibility; a new dimension of
   negotiation (really a sub-dimension of the feature dimension) can be
   added without the need for a new standards effort by the simple
   registration of a `feature tag'.  Feature registration is discussed
   in Section 8.


5  Variant descriptions

5.1 Syntax

   A variant can be described in a machine-readable way with a variant
   description.

       variant-description =
                  "{" <"> URI <"> source-quality *variant-attribute "}"

       source-quality = qvalue

       variant-attribute = "{" "type" media-type "}"
                         | "{" "charset" charset "}"
                         | "{" "language"  1#language-tag "}"
                         | "{" "length" 1*DIGIT "}"
                         | "{" "features" feature-list "}"
                         | "{" "description" quoted-string "}"
                         | extension-attribute

       extension-attribute = "{" extension-name extension-value "}"
       extension-name      = token
       extension-value     = *( token | quoted-string | LWS
                              | extension-specials )

       extension-specials  = <any element of tspecials except <"> and "}">

   Examples are

      {"paper.html.fr" 0.7 {type text/html} {language fr}}

      {"paper.html.tables"  0.9 {type text/html} {features tables}}

   The various attributes which can be present in a variant
   description are covered in the subsections below.  Each attribute
   may appear only once in a variant description.


5.2 URI

   The URI attribute gives the URI of the resource from which the
   variant can be retrieved with a GET request.  It can be absolute or
   relative to the Request-URI.  The variant resource may vary (on the
   Cookie request header, for example), but must not engage in
   transparent content negotiation itself.


5.3 Source-quality

   The source-quality attribute gives the quality of the variant, as a
   representation of the negotiable resource, when this variant is
   rendered with a perfect rendering engine on the best possible
   output medium.

   If the source-quality is less than 1, it often expresses a quality
   degradation caused by a lossy conversion to a particular data
   format.  For example, a picture originally in JPEG form would have
   a lower source quality when translated to the XBM format, and a
   much lower source quality when translated to an ASCII-art variant.
   Note however, that degradation is a function of the source; an
   original piece of ASCII-art may degrade in quality if it is
   captured in JPEG form.

   The source-quality could also represent a level of quality
   caused by skill of language translation, ability of a feature
   set to capture an artistic expression, or clarity of mathematical
   expression using a particular character set.

   It is important that content providers do not assign very low
   source quality values without good reason, as this would limit the
   ability of users to influence the negotiation process with their
   own preference settings.  The following table should be used as a
   guide when assigning source quality values:

       1.000       perfect representation
       0.999-0.900 near threshold of noticeable loss of quality
       0.899-0.700 noticeable, but acceptable quality reduction
       0.699-0.500 barely acceptable quality
       0.499-0.000 severely degraded quality

   Note that most meaningful values in this table are close to 1.
   This is due to the fact that quality factors are combined by
   multiplying them, not by adding them.

   When assigning source-quality values, content providers must not
   account for the size of the variant and its impact on transmission
   and rendering delays.  Source-quality values are assigned assuming
   instantaneous transmission and rendering.  This rule ensures that a
   future mechanism for bandwidth negotiation (see Appendix 19.5.2)
   can be cleanly added.


5.4 Type, language, length, and charset.

   The type, language, and length attributes of a variant description
   can carry the same information as their Content-* response header
   counterparts as defined in [1].  The counterpart of the charset
   attribute is the charset parameter which can be present in the
   Content-Type response header, for example

      Content-Type: text/html; charset=ISO-8859-4

   Though all of these attributes are optional, it is often desirable
   to include as many attributes as possible as this will increase the
   quality of the negotiation process.

   The length attribute (if present) must reflect the length of the
   variant alone, and not the sum of the lengths of the variant and
   any objects inlined or embedded by the variant.

      Note: A server is not required to maintain a one-to-one
      correspondence between attributes in the description of a
      variant and the Content-* headers in responses containing that
      variant.  Examples: If the variant description contains a
      language attribute, the response does not necessarily have to
      contain a Content-Language header. If a Content-Language
      header is present, it does not have to contain an exact copy of
      the information in the language attribute.


5.5 Features

   The features attribute specifies how the presence or absence of
   particular features in the user agent affects the overall quality of
   the variant.  This attribute is covered in Section 6.4.


5.6 Description

   The description attribute gives a textual description of the
   variant.  It can be included if the URI and normal attributes of a
   variant are considered too opaque to allow interpretation by the
   user.  If a user agent is showing a menu of available variants
   compiled from a variant list, and if a variant has a description
   attribute, the user agent should show the description attribute of
   the variant instead of showing the normal attributes of the
   variant.


5.7 Extension-attribute

   The extension-attribute allows future specifications to
   incrementally define new dimensions of negotiation, and eases
   content negotiation experiments under HTTP/1.1.  In experimental
   situations, servers must only generate extension-attributes whose
   names start with "x-".  User agents should ignore all extension
   attributes they do not recognize.  Proxies must not run the network
   negotiation algorithm if an unknown extension attribute is present
   in the variant list.


6  Feature negotiation

   Feature negotiation has been introduced in Section 4.7.  This
   section defines the feature negotiation mechanism.  Section 7
   contains examples of its use.


6.1 Feature tags

   A feature tag (ftag) identifies a capability of a user agent or a
   preference of a user.  A feature is said to be `present' in a user
   agent if the corresponding capability is implemented, or if the
   user has expressed corresponding preference.

       ftag = 1*<any CHAR except CTLs or tspecials or "!">

       tspecials      = "(" | ")" | "<" | ">" | "@"
                      | "," | ";" | ":" | "\" | <">
                      | "/" | "[" | "]" | "?" | "="
                      | "{" | "}" | SP | HT

        (tspecials definition reproduced from [1])

   Examples are

        tables, fonts, blebber, wolx, screenwidth, colordepth

   An example of the use of feature tags in a variant description is:

     {"index.html" 1.0 {type text/html} {features tables frames}}

   Feature-tags are case-insensitive.  The definition of a feature tag
   may state that a feature tag, if present, can have associated with
   it one or more values which reflect a particular capability or
   preference.  For example, a feature tag `paper' could be present
   with the values `A4' and `A5'.

   Note that context may determine whether a feature tag expresses a
   capability or a preference.  The `textonly' tag is naturally
   present for a text-only browser, but the user of a graphical
   browser could set the tag to be present if text-only content is
   preferred to graphical content.

   As feature registration will be an ongoing process, it is generally
   not possible for a user agent to know the meaning of all feature
   tags it can possibly encounter in a variant description.  A user
   agent should treat all features with tags unknown to it as absent.


6.2 Accept-Features header

   The Accept-Features request header can be used by a client to give
   information about the presence or absence of certain features.

       Accept-Features = "Accept-Features" ":"
                   #( feature-expr *( ";" feature-extension ) )

       feature-expr = [ "!" ] ftag [ "=" tag-value ]
                    | ftag "=" "{" tag-value "}"
                    | ftag "<=" number
                    | ftag "=" "<" numeric-range ">"
                    | "*"

       tag-value  = token | quoted-string

       number = 1*DIGIT

       numeric-range = [ number ] "-" [ number ]

       feature-extension = token [ "=" ( token | quoted-string ) ]

       [##Note: an earlier version of this specification did not allow
       quoted strings as values.  Quoted strings were added because
       they make it easier to map existing negotiation dimensions and
       formats onto the feature negotiation framework.##]

   Tag values must be compared case-insensitively, and a token value
   XYZ is equal to a quoted-string value "XYZ".  No feature extensions
   are defined in this specification.  An example is:

       Accept-Features: blex, !blebber, colordepth<=5, !screenwidth,
                  UA-media={stationary}, paper=a4, !paper="a0",
                  x_version=<100-205>, *

   The different feature expressions have the following meaning:

      ftag       ftag is present

      !ftag      ftag is absent

      ftag=V     ftag is present with the value V

      !ftag=V    ftag is present, but not with the value V

      ftag={V}   ftag is present with the value V, and not with any
                 other values

      ftag<=N    ftag is present with the numeric values from 0 up to
                 and including N, and not with any other values

      ftag=<N-M> ftag is present with the numeric values from N up to
                 and including M, and not with any other values.  If N
                 is missing, the lower bound is 0.  If M is missing,
                 the upper bound is infinity.

      *          makes true all feature predicates (Section 6.3) which
                 were not assigned truth values by other elements of
                 the header

      [##Question to be resolved: is there an alternative to the
      ftag=<N-M> notation which looks nicer?  We can't use ftag=N-M
      because it would be syntactically ambiguous with ftag=V: N-M is
      also a HTTP token value.  Maybe tag@N-M is a nicer notation.##]

   Absence of the Accept-Features header in a request is equivalent to
   the inclusion of

       Accept-Features: *

   [##Note: One possible use of the feature-extensions would be the
   introduction of quality values on feature tags, see Appendix
   19.3C##]


6.3 Feature predicates

   Feature predicates are used in the features attribute of a variant
   description.

      fpred = [ "!" ] ftag [ "=" tag-value ]
            | ftag "=" "<" numeric-range ">"

   Examples of feature predicates are

      blebber, !blebber, paper=a4, colordepth=5, !blex=54,
      dpi=<300-599>, colordepth=<24->

   The network negotiation algorithm can compute the truth value of a
   feature predicate by using the contents of the Accept-Features
   header of the current request.

   The truth value of a predicate is assigned as follows, depending on
   its form, by using the information in the Accept-Features
   header:

      ftag       true if the feature is known to be present

                 false if the feature is known to be absent

      !ftag      true if the feature is known to be absent

                 false if the feature is known to be present

      ftag=V     true if the feature is known to be present with
                 the value V,

                 false if the feature is known not to be present with
                 the value V

      !ftag=V    true if the feature is known to be present, but known
                 not to be present with the value V,

                 false if the feature is known to be absent or present
                 with the value V

      ftag=<N-M> true if the feature is known to be present with some
                 numeric values, while the highest value with which it
                 is present is known and in the range N-M,

                 false if the feature is known to be absent, or if it
                 is known to be present with some numeric values,
                 while the highest value with which it is present is
                 known and not in the range N-M.

                 If N is missing, the lower bound is 0.  If M is
                 missing, the upper bound is infinity.

   If the information in the Accept-Features header does not provide
   sufficient knowledge to assign a value to a predicate using the
   above rules, then the value is true if there is a "*" in the
   Accept-Features header, false otherwise.

   As an example, the header

       Accept-Features: blex, !blebber, colordepth<=5, !screenwidth,
                  UA-media={stationary}, paper=a4, !paper="a0",
                  x_version=<100-205>, *

   makes the following predicates true:

       blex, colordepth=4, !colordepth=6, colordepth, !screenwidth,
       UA-media=stationary, !UA-media=screen, paper=a4, !paper=a0,
       colordepth=<4-6>, !colordepth=<10->, x_version="101",
       blex=wox, !blex=wox,  paper=a5,
       frtnbf, !frtnbf, frtnbf=4, !frtnbf=4, frtnbf=<1-42>

   and makes the following predicates false:

       !blex, blebber, colordepth=6, colordepth=foo, !colordepth,
       screenwidth, screenwidth=640, !screenwidth=640, x_version=99,
       UA-media=screen, paper=a0


6.4 Features attribute

   The features attribute

            "{" "features" feature-list "}"

   is used in a variant description to specify how the presence or
   absence of particular features in the user agent affects the overall
   quality of the variant.

       feature-list = 1%feature-list-element

       feature-list-element = ( fpred | fpred-bag )
                              [ ":" true-improvement ]
                              [ "/" false-degradation  ]

       fpred-bag = "[" 1%fpred "]"

       true-improvement   =  1*3DIGIT [ "." 0*3DIGIT ]
       false-degradation  =  1*3DIGIT [ "." 0*3DIGIT ]

   Examples are:

       {features !textonly [blebber !wolx] colordepth=3:0.7 }

       {features !blink/0.5 background:1.5  [blebber !wolx]:1.4/0.8 }

   The default value for the true-improvement is 1.  The default value
   for the false-degradation is 0, or 1 if a true-improvement value is
   given.

   The network negotiation algorithm can compute the quality
   degradation factor associated with the features attribute by
   multiplying all quality degradation factors of the elements of the
   feature-list.  Note that the result can be a factor greater than 1.

   A feature list element yields its true-improvement factor if the
   corresponding feature predicate is true, or if at least one element
   of the corresponding fpred-bag is true. The element yields its
   false-degradation factor otherwise.

   [##Note: The attribute grammar above was designed to be parseable
   with simple non-recursive parsers.  The true-improvement construct
   does not add expressive power in a theoretical sense, but does make
   the (automatic) construction of variant lists more straightforward
   in many cases.##]

   [##Aside for mathematicians: note the similarity between the
   feature-list

            [f1 !f2] [!f1 f3]

   and the conjunctive normal form of a boolean expression.  This
   similarity implies good things about the expressive power of the
   features attribute.##]


7  Feature negotiation examples

   This section contains examples of the use of feature tags in
   variant descriptions.

   [##Note: A future version of this document will probably revise and
   extend this section.##]


7.1 Use of feature tags

   Feature tags can be used in variant lists to express the quality
   degradation associated with the presence or absence of certain
   features.  One example is

     {"index.html.plain" 0.7 },
     {"index.html"       1.0 {features tables frames}}

   Here, the "{features tables frames}" part expresses that index.html
   uses the features tagged as tables and frames.  If these features
   are absent, the overall quality of index.html degrades to 0.
   Another example is

     {"home.graphics" 1.0 {features !textonly}},
     {"home.textonly" 0.7 }

   where the "{features !textonly}" part expresses that home.graphics
   requires the absence of the textonly feature.  If the feature is
   present, the overall quality of home.graphics degrades to 0.

   The absence of a feature need not always degrade the overall quality
   to 0.  In the example

     {"x.html.1" 1.0 {features fonts/0.7}}

   the absence of the fonts feature degrades the quality with a factor
   of 0.7.  "fonts/0.7" can be pronounced as "fonts, or a degradation of
   0.7".  In the example

     {"x.html.2" 0.5 {features fonts:1.5}}

   the presence of the fonts feature improves the overall quality with a
   factor of 1.5. If the fonts feature is absent, the overall quality is
   not affected.  "fonts:1.5" can be pronounced as "fonts improves with
   1.5".  Finally, in the example

      {"y.html" 1.0 {features [blebber wolx] }}

   The "[blebber wolx]" expresses that y.html requires the presence of
   the blebber feature or the wolx feature.  This construct can be
   used in a number of cases:

     1. blebber and wolx actually tag the same feature, but they
        were registered by different people, and some browsers say
        they support blebber while others say they support wolx.

     2. blebber and wolx are HTML tags of different vendors which
        implement the same functionality, and which were used
        together in y.html without interference.

     3. blebber and wolx are HTML tags of different vendors which
        implement the same functionality, and y.html uses the tags in
        a conditional HTML construct.

     4. blebber is a complicated HTML tag with only a sketchy
        definition, implemented by one browser vendor, and wolx
        indicates implementation of a well-defined subset of the blebber
        tag by some other vendor(s).  y.html uses only this well-defined
        subset.


7.2 Use of numeric feature tags

   As an example of negotiation in a numeric area, the following
   variant list describes four variants with title graphics designed
   for increasing screen widths:

     {"home.pda"    1.0 {features pagewidth=<-199> }},
     {"home.narrow" 1.0 {features pagewidth=<200-599> }},
     {"home.normal" 1.0 {features pagewidth=<600-999> }},
     {"home.wide"   1.0 {features pagewidth=<1000-> }}

   The accept header

     Accept-Features: pagewidth<=250, *

   which indicates that the pagewidth tag is present with the
   numeric values from 0 up to and including 250, makes the following
   feature predicates false:

     pagewidth=<-199>  pagewidth=<600-999>  pagewidth=<1000->

   and thereby degrades the overall quality of all above variants
   except home.narrow to 0.  The above variant list has a serious
   problem, however: a user agent not supporting pagewidth
   negotiation is given no guidance on which variant to select.  The
   header

     Accept-Features: !pagewidth, *

   degrades the overall quality of all above variants to 0.

   A variant list which solves this problem is

     {"home.pda"    1.0 {features pagewidth=<-199> }},
     {"home.narrow" 1.0 {features pagewidth=<200-599> }},
     {"home.normal" 0.99 },
     {"home.wide"   1.0 {features pagewidth=<1000-> }}

   With this list, a user agent which does not support the pagewidth
   feature will always select the home.normal variant.  Another
   solution would be to describe the normal variant as:

     {"home.normal" 1.0 {features [pagewidth=<600-999> !pagewidth] }}


8  Feature tag registration

   Feature tag registration procedures are defined in [4].  Note that
   the feature tag names in the examples of this specification do not
   in general reflect the tag naming schemes proposed in [4].


8.1 Evolution of feature tags

   The feature negotiation mechanism is designed not to break if 1000+
   features, partially overlapping, are registered.  This allows for
   feature tag creation to be an evolutionary process, in which many
   tags are created while only a few `survive' to become generally
   used.

   As the development of new web content formats is currently in an
   evolutionary phase, rather than a standardization phase, it is
   thought that this evolutionary approach to feature tag creation has
   the best chances of keeping up with new developments.


8.2 Core set of feature tags

   [##In order to jump-start feature negotiation, it seems useful to
   define a `core set' of feature tags in a separate document.  These
   feature tags would cover the areas of negotiation which are
   currently well understood, like negotiation on currently stable
   HTML extensions.  The transparent content negotiation specification
   could then require (or strongly encourage) that user agents
   implementing transparent content negotiation must recognize (not
   support!) all tags in the core set.

   The core set will be proposed in a non-standards track document
   which will be completed along with this document.##]


8.3 Feature tag design

   When designing a new feature tag, it is important to take into
   account that existing user agents which do not recognize the new
   tag will treat the feature as absent.  In general, a new feature
   tag needs to be designed in such a way that absence of the tag is
   the default case which reflects current practice.  If this design
   principle is ignored, the resulting feature tag will generally be
   unusable.

   As an example, one could try to support negotiation between
   monochrome and color content by introducing a `color' feature tag,
   the presence of which would indicate the capability to display
   color graphics.  However, if this new tag is used in a variant
   list, for example

      {"rainbow.gif"      1.0 {features color} }
      {"rainbow.mono.gif" 0.6 {features !color}}

   then existing user agents, which would not recognize the color tag,
   would all display the monochrome rainbow.  The color tag is
   therefore unusable in situations where optimal results for existing
   user agents are desired.  To provide for negotiation in this area,
   one must introduce a `monochrome' feature tag; its presence
   indicates that the user agent can only render (or the user prefers
   to view) monochrome graphics.


9  Content negotiation response codes and headers

   This specification adds one new response code to those defined in
   HTTP/1.1 [1].  This is a 5xx (Server Error) class response.  It
   also adds six new headers.


9.1 506 Variant Also Negotiates

   The server has an internal configuration error; the chosen variant
   resource is configured to engage in transparent content negotiation
   itself, and is therefore not a proper end point in the negotiation
   process.


9.2 Accept-Features

  This request header was defined in Section 6.2.


9.3 Content-Features

   The Content-Features response header can be used by a server to
   indicate how the presence or absence of particular features in the
   user agent affects the overall quality of the response.

       Content-Features = "Content-Features" ":" feature-list


9.4 Alternates

   The Alternates response header is used to convey the list of
   variants bound to a negotiable resource in a response.  It can also
   contain other directives for the content negotiation process.

       Alternates = "Alternates" ":" 1#( variant-description
                                       | alternates-directive )

       alternates-directive = ( "min-q" "=" short-float )
                            | extension-alternates-directive

       delta-seconds  = 1*DIGIT

       short-float = 1*3DIGIT [ "." 0*3DIGIT ]

       extension-alternates-directive =
                              token [ "=" ( token | quoted-string ) ]

   An example is

     Alternates: {"paper.html.en" 0.9 {type text/html} {language en}},
                 {"paper.html.fr" 0.7 {type text/html} {language fr}},
                 {"paper.ps.en"   1.0 {type application/postscript}
                     {language en}},
                 min-q=0.4

   Any relative URI specified in a variant-description field is
   relative to the request-URI.  The min-q directive, if present, can
   influence decisions made by the network negotiation algorithm on
   behalf of the origin server.  Clients should ignore all
   extension-alternates-directives they do not understand.

   A variant list may contain multiple differing descriptions of the
   same variant.  This can be convenient if the variant uses
   conditional rendering constructs, or if the variant resource
   returns multiple representations using a multipart media type.


9.6 Negotiate

   The Negotiate request header can contain directives for any content
   negotiation process initiated by the request.  Presence of the
   `trans' directive in this header indicates that the user agent
   supports transparent content negotiation for the current request.

      Negotiate = "Negotiate" ":" 1#negotiate-directive

      negotiate-directive = "trans" | negotiate-extension

      negotiate-extension = token [ "=" token ]

   An example is

      Negotiate: trans

   Servers should ignore all negotiate-directives they do not
   understand.  The presence of a `trans' directive is significant in
   calculations performed by the network negotiation algorithm.

   [#Note: The name `trans' is a compromise between the long
   `transparent' and the short `t'.  The HTTP-WG is divided on the
   naming issue, and `trans' was chosen to minimize overall
   unhappiness.  The WG will only be truly happy if a shorter synonym
   of `negotiate' can be found.##]


9.7 Variant-Vary

   The Variant-Vary response header can be used in a list response to
   record any vary information which applies to the variant data
   contained in the response, rather than to the response as a whole.

      Variant-Vary  = "Variant-Vary" ":" ( "*" | 1#field-name )

   Use of the Variant-Vary header is discussed in Section 10.2.


9B Cache validators

   To allow for correct and efficient caching and revalidation of
   negotiated responses, this specification extends the caching model
   of HTTP/1.1 [1] in various ways.

   Under the rules in this specification, the maximal age (time since
   last revalidation) of a variant list bound to a negotiable
   resource, as received from a cache in an Alternates header, is the
   maximum of

      1. the freshness lifetimes (max-age values) in the responses
         from the negotiable resource itself, and

      2. the freshness lifetimes (max-age values) of the variant
         resources of the negotiable resource which are also neighbors
         of the negotiable resource.

   If no freshness lifetimes are assigned by the origin server, the
   maximal age of a variant list is the maximum of the freshness
   lifetime values which were heuristically assigned by the cache.


9B.1 Variant list validators

   A variant list validator is an opaque value which acts as the cache
   validator of a variant list bound to a negotiable resource.

      variant-list-validator = <quoted-string not containing any ";">

   If two responses contain the same variant list validator, a cache
   can treat the Alternates headers in these responses as equivalent
   (though the headers themselves need not be identical).


9B.2 Structured entity tags

   A structured entity tag consists of a normal entity tag of which
   the opaque string is extended with a semicolon followed by a
   variant list validator:

        normal      |  variant list  |   structured
        entity tag  |  validator     |   entity tag
       -------------+----------------+-----------------
         "etag"     |     "vlv"      |   "etag;vlv"
        W/"etag"    |     "vlv"      |  W/"etag;vlv"

   Note that a structured entity tag is itself also an entity tag.
   The structured nature of the tag allows caching proxies capable of
   transparent content negotiation to perform certain optimizations.
   Examples of structured entity tags are:

      "xyzzy;1234"  W/"xyzzy;1234"  "gonkxxxx;1234"  "a;b;c;;1234"

   In the last example, the normal entity tag is "a;b;c;" and the
   variant list validator is "1234".

   If a transparently negotiated response includes an entity tag, it
   must be a structured entity tag.  The variant list validator in the
   structured tag must act as a validator for the variant list
   contained in the Alternates header.  The normal entity tag in the
   structured tag must act as a validator of the entity body in the
   response and of all entity headers except Alternates.


9B.3 Assigning entity tags to variants

   To ensure the correct revalidation of transparently negotiated
   responses, origin servers should maintain some uniqueness
   requirements when assigning entity tags to variant resources.

   First, two different variant resources of the same negotiable
   resource, which are both neighbors of this negotiable resource, may
   never use the same (normal) entity tag, even not at different
   times, unless this tag labels exactly the same entity on all
   occasions.

   Second, the server should ensure that, if a normal entity tag of a
   neighboring variant resource is extended to a structured tag, this
   structured tag never equals a different normal entity tag used by
   another variant or at another time.  An easy way to ensure this is
   by never including a semicolon in a normal tag.


10 Content negotiation responses

   If a request on a transparently negotiated resource yields a
   response with a 2xx status code or any 3xx status code except 304,
   this response must always be either a list response, a choice
   response, or an ad hoc response.  These responses always include
   the Alternates header bound to the negotiable resource.

   Transparently negotiated responses with other status codes may also
   include an Alternates header, if this is allowed by the HTTP/1.1
   specification [1].  Note that HTTP/1.1 does not allow an Alternates
   header in a 304 (Not Modified) response.  When generating a
   response from a resource which does not support transparent content
   negotiation, a server must never include an Alternates header.

   A list response always has the 300 (Multiple Choices) response
   code.  A choice response never has the 300 code, and always has a
   Content-Location header.  An ad hoc response never has the 300
   code, and never has a Content-Location header.

10.1 List response

   A list response has the 300 response status code.  It must contain
   (besides the normal headers required by HTTP) the Alternates header
   bound to the negotiable resource, a Vary header and (unless it was
   a HEAD request) an entity body which allows the user to manually
   select the best variant.  It is generated as a response to a user
   agent capable of transparent content negotiation if the server does
   not (perhaps cannot) choose a particular best variant for the
   request.

   An example of a list response is

     HTTP/1.1 300 Multiple Choices
     Date: Tue, 11 Jun 1996 20:02:21 GMT
     Alternates: {"paper.html.en" 0.9 {type text/html} {language en}},
                 {"paper.html.fr" 0.7 {type text/html} {language fr}},
                 {"paper.ps.en"   1.0 {type application/postscript}
                     {language en}}
     Vary: negotiate, accept, accept-language
     ETag: "blah;1234"
     Cache-control: max-age=86400
     Content-Type: text/html
     Content-Length: 227

     <h2>Multiple Choices:</h2>
     <ul>
     <li><a href=paper.html.en>HTML, English version</a>
     <li><a href=paper.html.fr>HTML, French version</a>
     <li><a href=paper.ps.en>Postscript, English version</a>
     </ul>

   The Vary header in the response should ensure correct handling by
   HTTP/1.1 caching proxies not capable of transparent content
   negotiation.  This header can either be

        Vary: *

   or a more elaborate header; see Section 10.5.1.

   Only the origin server may construct list responses.  List
   responses are cacheable unless indicated otherwise.

   According to the HTTP/1.1 specification [1], a user agent not
   capable of transparent content negotiation will, when receiving a
   list response, display the entity body included in the response.
   If the response contains a Location header, however, the user agent
   may automatically redirect to this location.

   The handling of list responses by clients supporting transparent
   content negotiation is described in Sections 12.1 and 14.

      Note: Some existing versions of HTTP/1.0 clients are known to
      silently ignore list responses, instead of handling them
      according to the HTTP/1.0 specification [3].  Servers should
      therefore be careful in sending list responses to
      non-negotiating HTTP/1.0 user agents, and in making these
      responses cacheable.


10.2 Choice response

   A choice response merges a normal HTTP response from the chosen
   variant, a Content-Location header giving the location of the
   variant, and the Alternates headers bound to the negotiable
   resource.  Depending on the response code, a choice response is
   cacheable unless indicated otherwise.

   Origin servers and proxy caches must construct choice responses
   with the following algorithm (or any other algorithm which gives
   equal results).

   In this algorithm, `the current Alternates header' refers to the
   Alternates header containing the variant list which was used to
   choose the best variant, and `the current variant list validator'
   refers to the validator of this list.  Section 10.3 specifies how
   these two things can be obtained by a proxy cache.

   The algorithm consists of five steps.

     1. Construct a HTTP request message on the best variant resource
        by rewriting the request-URI and Host header (if appropriate)
        of the received request message on the negotiable resource.

     2. Generate a valid HTTP response message, but not one with the
        304 (Not Modified) code, for the request message constructed
        in step 1.

        In a proxy cache, the response can be obtained from cache
        memory, or by passing the constructed HTTP request towards the
        origin server.  If the request is passed on, the proxy may
        add, modify, or delete If-None-Match and If-Range headers to
        optimize the transaction with the upstream server.

     3. Check for an origin server configuration error. If the HTTP
        response message generated in step 2 contains an Alternates
        header, a Content-Location header, or has the 300 status code,
        then the best variant resource is not a proper end point in
        the negotiation process, and a 506 (Variant Also Negotiates)
        error response message should be generated instead of going to
        step 4.

     4. Add a number of headers to the HTTP response message generated
        in step 2.

        a. Add a Content-Location header giving the location of the
           chosen variant.

               Note: According to the HTTP/1.1 specification [1], if
               the Content-Location header contains a relative URI,
               this URI is relative to the URI in the Content-Base
               header, if present.

        b. If any Vary headers are present in the response message
           from step 2, add, for every Vary header, a Variant-Vary
           header with a copy of the contents of this Vary header.

        c. Add the current Alternates header.

        d. Add a Vary header to ensure correct handling by HTTP/1.1
           caching proxies not capable of transparent content
           negotiation.  This header can either be

               Vary: *

           or a more elaborate header, see Section 10.5.

        e. To ensure compatibility with HTTP/1.0 caching proxies which
           do not recognize the Vary header, an Expires header with a
           date in the past may be added. See Appendix 19.1 for more
           information.

        f. If an ETag header is present in the response message from
           step 2, then extend the entity tag in that header with the
           current variant list validator, as specified in Section
           9B.2.

        g. Only in proxy caches: set the Age header of the response to

              max( variant_age , alternates_age )

           where variant_age is the age of the best variant response
           obtained in step 2, calculated according to the rules in
           the HTTP/1.1 specification [1], and alternates_age is the
           age of the Alternates header added in step c, calculated
           according to the rules in Section 10.3.

     5. The response may be shortened to a 304 (Not Modified) or 206
        (Partial Content) response if an If-No-Match or If-Range
        header in the original request message allows it according to
        the HTTP/1.1 specification [1].

           Note: In this algorithm, a proxy cache always generates a
           full response message in step 2.  In practice, proxy caches
           can use a more efficient algorithm which produces equal
           results, but which avoids expanding 304 responses from
           upstream servers to full responses in step 2 if these
           responses will be shortened again in step 5.  By
           shortcutting the expansion, a proxy cache can sometimes
           generate a legal 304 response even if it has not cached the
           variant data itself.

   An example of a choice response is:

     HTTP/1.1 200 OK
     Date: Tue, 11 Jun 1996 20:05:31 GMT
     Content-Type: text/html
     Last-Modified: Mon, 10 Jun 1996 10:01:14 GMT
     Content-Length: 5327
     Cache-control: max-age=604800
     Content-Location: paper.html.en
     Alternates: {"paper.html.en" 0.9 {type text/html} {language en}},
                 {"paper.html.fr" 0.7 {type text/html} {language fr}},
                 {"paper.ps.en"   1.0 {type application/postscript}
                     {language en}}
     Etag: "gonkyyyy;1234"
     Vary: negotiate, accept, accept-language
     Expires: Thu, 01 Jan 1980 00:00:00 GMT

     <title>A paper about ....

   An example of forwarding by a proxy cache: if a proxy receives the
   request

     GET /paper HTTP/1.1
     Host: x.org
     User-Agent: WuxtaWeb/2.4
     Negotiate: trans
     Accept: text/html, *
     Accept-Language: en
     If-None-Match: "gonkyyyy;1234", W/"a;b;1234"

   and if it can reuse a cached variant list with the validator
   "1234", taken from a cached response with an age of 8000 seconds,
   to choose paper.html.en as the best variant, then the proxy can
   pass on the request

     GET /paper.html.en HTTP/1.1
     Host: x.org
     User-Agent: WuxtaWeb/2.4
     Negotiate: trans
     Accept: text/html, *
     Accept-Language: en
     If-None-Match: "gonkyyyy", W/"a;b"
     Via: 1.1 fred

   to an upstream server.  On receipt of the response

     HTTP/1.1 304 Not Modified
     Date: Tue, 11 Jun 1996 20:05:31 GMT
     Etag: "gonkyyyy"

   from the upstream server, it can return

     HTTP/1.1 304 Not Modified
     Date: Tue, 11 Jun 1996 20:05:31 GMT
     Etag: "gonkyyyy;1234"
     Content-Location: paper.html.en
     Vary: negotiate, accept, accept-language
     Expires: Thu, 01 Jan 1980 00:00:00 GMT
     Via: 1.1 fred
     Age: 8000

   to its own client.


10.2B Ad hoc response

   An ad hoc response never has the 300 response status code and never
   has a Content-Location header.  It must contain the Alternates
   header bound to the negotiable resource, and a Vary header if the
   response is cacheable.  It may be generated by an origin server,
   usually as a response to a non-negotiating user agent, if the
   server cannot or does not want to send a list or choice response.

   The Vary header in the response should ensure correct handling by
   HTTP/1.1 caching proxies not capable of transparent content
   negotiation.  This header can either be

        Vary: *

   or a more elaborate header, see Section 10.5.1.  Ad hoc responses
   are cacheable unless indicated otherwise.  The handling of ad hoc
   responses by clients supporting transparent content negotiation is
   described in Sections 12.1 and 14.

   An example of an ad hoc response is:

     HTTP/1.1 200 OK
     Date: Tue, 11 Jun 1996 20:02:26 GMT
     Alternates: {"paper.html.en" 0.9 {type text/html} {language en}},
                 {"paper.html.fr" 0.7 {type text/html} {language fr}},
                 {"paper.ps.en"   1.0 {type application/postscript}
                     {language en}}
     Vary: negotiate, accept, accept-language
     Etag: "gonkzzzz;1234"
     Cache-control: max-age=86400
     Content-Type: text/html
     Content-Length: 227

     <h2>Multiple Choices:</h2>
     <ul>
     <li><a href=paper.html.en>HTML, English version</a>
     <li><a href=paper.html.fr>HTML, French version</a>
     <li><a href=paper.ps.en>Postscript, English version</a>
     </ul>

   Another example is

     HTTP/1.1 302 Moved Temporarily
     Date: Tue, 11 Jun 1996 20:02:28 GMT
     Alternates: {"paper.html.en" 0.9 {type text/html} {language en}},
                 {"paper.html.fr" 0.7 {type text/html} {language fr}},
                 {"paper.ps.en"   1.0 {type application/postscript}
                     {language en}}
     Location: paper.html.en
     Content-Type: text/html
     Content-Length: 59

     This document is available <a href=paper.html.en>here</a>.


10.3 Reusing the Alternates header

   If a proxy cache has available a negotiated response which is
   cacheable, fresh, and has an ETag header, then it may extract the
   Alternates header and associated variant list validator from the
   response, and reuse them (without unnecessary delay) to negotiate
   on behalf of the user agent or origin server (Section 14) or to
   construct a choice response (Section 10.2).  The age of the
   extracted Alternates header is the age of the response from which
   it is extracted, calculated according to the rules in the HTTP/1.1
   specification [1].


10.4 Extracting a normal response from a choice response

   If a proxy receives a choice response, it may extract and cache the
   normal HTTP response contained therein.  The normal response can be
   extracted by taking a copy of the choice response and then deleting
   the Content-Location, Alternates, and Vary headers, renaming any
   Variant-Vary headers to Vary headers, and shortening the structured
   entity tag in any ETag header to a normal entity tag.

   This normal response may then be cached (as a HTTP response to the
   variant request as constructed in step 1. of Section 10.2) and
   reused to answer future direct requests on the variant resource
   according to the rules in the HTTP/1.1 specification [1].  This
   caching of extracted responses can increase overall efficiency with
   up to a factor 2.

   For security reasons (see Section 15.2), an extracted normal
   response may only be cached if the negotiable resource and the
   variant resource are neighbors.  If they are not neighbors, the
   proxy should reject the choice response as a probable spoofing
   attempt and pass on a 502 (bad gateway) error response instead.


10.5 Elaborate Vary headers

   If a HTTP/1.1 [1] server can generate varying responses for a
   request on a single resource, then the server must include a Vary
   header in these responses if they are cacheable.  This Vary header
   is a signal to HTTP/1.1 caches that something special is going on.
   It prevents the caches from returning the currently chosen response
   to every future request on the resource.

   Servers engaging in transparent content negotiation will generate
   varying responses.  Therefore, cacheable list, choice, and ad hoc
   responses must always include a Vary header.

   The most simple Vary header which can be included is

       Vary: *

   This header leaves the way in which the response is selected by the
   server completely unspecified.

   A more elaborate Vary header can be used to allow for certain
   optimizations in HTTP/1.1 caches which are not capable of
   transparent content negotiation, but which do cache multiple
   variant responses for one resource.  Such a more elaborate Vary
   header lists all request headers which can be used by the server
   when selecting a response for a request on the resource.

10.5.1 Construction of an elaborate Vary header

   Origin servers can construct a more elaborate Vary header in the
   following way.  First, start with the header

       Vary: negotiate

   `negotiate' is always included because servers using the network
   negotiation algorithm will sometimes use the information in the
   Negotiate header when choosing between a list, choice, or ad-hoc
   response.

   Then, if any of the following attributes is present in any variant
   description in the Alternates header, add the corresponding header
   name to the Vary header

         attribute  |   header name to add
         -----------+---------------------
          type      |   accept
          charset   |   accept-charset
          language  |   accept-language
          features  |   accept-features

   The Vary header constructed in this way specifies the response
   variation which can be caused by the use of the network negotiation
   algorithm in proxies.  If the origin server uses a custom
   negotiation algorithm which takes additional headers (for example
   User-Agent) into account, these should also be added to the Vary
   header.

10.5.2 Caching of an elaborate Vary header

   Proxy caches cannot construct elaborate Vary headers by themselves,
   because they may not assume that the origin server uses the
   standard network negotiation algorithm to vary its responses.
   However, when extracting an Alternates header from a response
   (Section 10.3) caches may also extract the Vary header in the
   response, and reuse it along with the Alternates header.  A clean
   Vary header can however only be extracted if the variant does not
   vary itself, i.e. if a Variant-Vary header is absent.


10.6 Negotiation on content encoding

   Negotiation on the content encoding of a response is orthogonal to
   transparent content negotiation.  The rules for when a content
   encoding may be applied are the same as in HTTP/1.1: servers may
   content-encode responses that are the result of transparent content
   negotiation whenever an Accept-Encoding header in the request
   allows it.  When negotiating on content encoding, servers must add
   the accept-encoding header name to the Vary header of the response,
   or add `Vary: *', if the response is cacheable.

   Servers should always be able to provide unencoded versions of
   every transparently negotiated response.  This means in particular
   that every variant in the variant list must at least be available
   in an unencoded form.

   Like HTTP/1.1, this specification allows proxies to encode or
   decode relayed or cached responses on the fly: the response still
   contains the same variant as far as transparent content negotiation
   is concerned.  Note that HTTP/1.1 requires proxies to add a Warning
   header if the encoding of a response is changed.


11 The network negotiation algorithm

   The network negotiation algorithm is a standardized algorithm by
   which a party in the negotiation process can make a choice on
   behalf of another party.  For example, a proxy could choose a
   particular variant on behalf of the user agent in order to speed up
   the negotiation process by cutting corners (see Section 4.4).

   General principles underlying the network negotiation algorithm are
   as follows.

    - When running on an origin server or proxy, the algorithm will
      usually only have partial user agent capabilities and
      preferences information as input.  Therefore, instead of a
      choice for a best variant, the algorithm can also yield an
      `insufficient data, cannot choose' result.

    - Decisions are always made on either on behalf of the user agent
      or on behalf of the origin server.  They are made on behalf of
      the user agent if the user agent indicates that it is capable of
      transparent content negotiation, and on behalf of the origin
      server otherwise.

    - By including an Alternates header in the response, the origin
      server gives upstream proxies the permission to perform actions
      prescribed by the network negotiation algorithm instead of
      actions prescribed by the plain HTTP specification.

    - It is expected that user agents will use the network negotiation
      algorithm by default, locally supplying their complete
      capabilities and preferences to the algorithm, but this is not
      required.  User agents may use customized negotiation algorithms
      to automatically choose a variant.  Such algorithms could for
      example account for cross-dependencies between dimensions of
      negotiation.

   Note that in this specification, `use of the network negotiation
   algorithm' means use of any algorithm which yields equal results.


11.1 Input

   The algorithm is always executed for a particular request on a
   particular transparently negotiable resource.  It takes the
   following information as input.

    1. The list of variants of the resource, as present in the
       Alternates header of the resource.

    2. Any "min-q" directive pertaining to the negotiation process, as
       specified by the Alternates header of the resource.

    3. (Partial) Information about capabilities and preferences of the
       user agent for this particular request, as given in the Accept
       headers of the request, or, if the algorithm runs on a user
       agent, as given by the internal capabilities and preferences
       database.

    4. A boolean indicating whether or not the user agent is capable
       of transparent content negotiation for this request.  (This can
       be determined by looking for a `trans' directive in a Negotiate
       header, see Section 9.6.)


11.2 Output

   As its output, the network negotiation algorithm and will yield the
   appropriate action to be performed.  Actions have labels `ACT_UA'
   for actions performed on behalf of the user agent, and `ACT_OS' for
   actions performed on behalf of the origin server.

11.2.1 Output for proxies

   If the algorithm runs on a proxy, there are four different possible
   results.  The first two results can be generated if the user agent
   is capable of transparent content negotiation:

      Choice_UA

        Choice on behalf of the user agent. The best variant may be
        retrieved and returned in a choice response.

      List_UA

        No choice possible, return the list.  The Accept headers do
        not contain sufficient information to make a choice on behalf
        of the user agent possible.  A list response should be
        returned, allowing the user agent to make the choice itself.

   The following results can be generated if the user agent is not
   capable of transparent content negotiation:

      Choice_OS

        Choice on behalf of the origin server.  The best variant may
        be retrieved and returned in a choice response.

      Forward_OS

        No choice possible, forward the request towards the origin
        server. The Accept headers do not contain sufficient
        information to make a choice on behalf of the origin server
        possible. The request should be forwarded towards the origin
        server so that it can choose the most appropriate action with
        a custom algorithm.

11.2.2 Output for origin servers

   If the algorithm runs on an origin server, all results above are
   possible.  The last result has a different interpretation:

      Forward_OS

        No choice possible, internally forward the request to a custom
        algorithm which chooses an appropriate action.

   Usually, the appropriate action is the sending of an ad hoc
   response.

11.2.3 Output for user agents

   If the algorithm runs on a user agent, only the results List_UA and
   Choice_UA are possible, with the following interpretations.

      Choice_UA

        Automatic choice. A particular variant X must be retrieved and
        displayed.

      List_UA

        No automatic choice possible, none of of the variants can be
        rendered according to the preferences and capabilities
        database.  An appropriate action like showing an error message
        with the list of variants to the user must be performed.


11.3 Computing the overall quality values

   As a first step in the network negotiation algorithm, the overall
   qualities of the individual variants in the list are computed.

   The overall quality Q of a variant is the value

      Q = qs * qt * qc * ql * qf

   where the factors qs, qt, qc, ql, and qf are determined as follows.

      qs Is the source quality factor in the variant description.

      qt The media type quality factor is 1 if there is no type
         attribute in the variant description, or if there is no
         Accept header in the request.  Otherwise, it is the quality
         assigned by the Accept header to the media type in the type
         attribute.

           Note: If a type is matched by none of the elements of an
           Accept header, the Accept header assigns the quality factor
           0 to that type.

      qc The charset quality factor is 1 if there is no charset
         attribute in the variant description, or if there is no
         Accept-Charset header in the request.  Otherwise, the charset
         quality factor is the quality assigned by the Accept-Charset
         header to charset in the charset attribute.

           Note: If a charset does not appear in an Accept-Charset
           header, the Accept-Charset header assigns the quality
           factor 0 to that charset.

      ql The language quality factor is 1 if there is no language
         attribute in the variant description, or if there is no
         Accept-Language header in the request.  Otherwise, the
         language quality factor is the highest quality factor
         assigned by the Accept-Language header to any one of the
         languages listed in the language attribute.

      qf The features quality factor is 1 if there is no features
         attribute in the variant description, or if there is no
         Accept-Features header in the request.  Otherwise, it is
         the feature negotiation quality degradation value for the
         features attribute, see Section 6.4.

   As an example, if a variant list contains the variant description

     {"paper.html.en" 0.7 {type text/html} {language fr}}

   and if the request contains the Accept headers

     Accept: text/html:q=1.0, */*:q=0.8
     Accept-Language: en;q=1.0, fr;q=0.5

   the network negotiation algorithm will compute an overall quality
   for the variant as follows:

     {"paper.html.en" 0.7 {type text/html} {language fr}}
                       |           |                 |
                       |           |                 |
                       V           V                 V
                      0.7   *     1.0        *      0.5  = 0.35

   With the above Accept headers, the complete variant list

     {"paper.html.en" 0.9 {type text/html} {language en}},
     {"paper.html.fr" 0.7 {type text/html} {language fr}},
     {"paper.ps.en"   1.0 {type application/postscript} {language en}}

   would yield the following computations:

                     qs  * qt  * qc  * ql  * qf  =  Q
                     ---   ---   ---   ---   ---   ----
      paper.html.en: 0.9 * 1.0 * 1.0 * 1.0 * 1.0 = 0.9
      paper.html.fr: 0.7 * 1.0 * 1.0 * 0.5 * 1.0 = 0.35
      paper.ps.en:   1.0 * 0.8 * 1.0 * 1.0 * 1.0 = 0.8


11.4 Definite and speculative quality values

   An overall quality value computed by an origin server or proxy can
   be either definite or speculative.

   An overall quality value for a variant is definite if it was
   computed without using any wildcard characters '*' in the Accept
   headers, and without the need to use the absence of a particular
   Accept header.  An overall quality value is speculative otherwise.

   As an example, in the previous section, the quality values of
   paper.html.en and paper.html.fr are definite, and the quality value
   of paper.ps.en is speculative because the type
   application/postscript was matched to the range */*.

   Definiteness can be defined more formally as follows.  An overall
   quality value Q is definite if the same quality value Q could be
   computed after the request message is changed in the following way:

       1. If an Accept, Accept-Charset, Accept-Language, or
          Accept-Features, header is missing from the request, add
          this header with an empty field.

       2. Delete any media ranges containing a wildcard character '*'
          from the Accept header.  Delete the language-range '*', if
          present, from the Accept-Language header.  Delete the
          wildcard '*', if present, from the Accept-Features header.

   As another example, the overall quality factor for the variant

     {"blah.html" 1 {language en-gb} {features blebber [x y]}}

   is 1 and definite with the Accept headers

     Accept-Language: en-gb, fr
     Accept-Features: blebber, x, !y, *

   and

     Accept-Language: en, fr
     Accept-Features: blebber, x, *

   The overall quality factor is still 1, but speculative, with the
   Accept headers

     Accept-language: en-gb, fr
     Accept-Features: blebber, !y, *

   and

     Accept-Language: fr, *
     Accept-Features: blebber, x, !y, *


11.5 Determining the result

   The best variant, as determined by the network negotiation
   algorithm, is the one variant with the highest overall quality
   value, or, if there are multiple variants which share the highest
   overall quality, the first variant in the list with this value.

11.5.1 Result for proxies and origin servers

   When running on a proxy or origin server, the result of the network
   negotiation algorithm is determined as follows.

   1. If the negotiable resource and the best variant resource are not
      neighbors, then the result is List_UA if the user agent is
      capable of transparent content negotiation, Forward_OS otherwise.
      This rule exists for security reasons: it prevents some forms of
      spoofing, see Section 15.2.

   2. If the above rule does not apply, then

      1. If the user agent is capable of transparent content
         negotiation, the result is Choice_UA if the best variant has
         an overall quality value greater than 0 which is also a
         definite quality value, and List_UA otherwise.

      2. If the user agent is not capable of transparent content
         negotiation, the result is Choice_OS if the best variant has
         a (definite or speculative) overall quality value which is
         greater than 0, or at least the value of the "min-q"
         directive if this directive is present in the Alternates
         header.  The result is Forward_OS otherwise.

11.5.2 Result for user agents

   When running on a user agent, the result of the network negotiation
   algorithm is Choice_UA if the best variant has an overall quality
   value greater than 0, and List_UA otherwise.


11.6 Construction of short requests

   Under normal HTTP semantics, the interpretation of the request
   header `Accept: */*' is

     `I accept all media types with a quality of 1.0'.

   One of the most important properties of the network negotiation
   algorithm is that this interpretation is different if the header is
   sent by a user agent which supports transparent content
   negotiation. In this case, the interpretation of the header is

     `I accept several media types, and assign quality factors from 0.0
     up to 1.0 to them.  If this information is insufficient to make a
     choice on my behalf, do not make a choice but send the list of
     variants'.

   According to the HTTP/1.1 specification [1], the complete absence
   of an Accept header from the request is to be interpreted in the
   same way.  This means that a user agent which supports transparent
   content negotiation can send a minimal request, without any Accept
   headers but with a Negotiate header, by default:

      GET /paper HTTP/1.1
      Host: x.org
      User-Agent: WuxtaWeb/2.4
      Negotiate:

   The sending of this minimal request will never have an adverse
   effect on the quality of the transparent negotiation process: if
   the requested resource happens to be transparently negotiated, the
   user agent will always get a list response, so that it can choose
   the best variant itself.  Only the speed of the negotiation process
   can be affected, because servers will no longer be able to cut
   corners by choosing on behalf of the user agent.

   If it is found that the requests to a particular origin server
   often return a less efficient list response, the user agent can
   dynamically introduce short Accept headers in its future requests
   to that server, for example

      GET /paper HTTP/1.1
      Host: x.org
      User-Agent: WuxtaWeb/2.4
      Negotiate:
      Accept: text/html, application/postscript:q=0.8, */*
      Accept-language: en, fr;q=0.5

   This will increase the chance that the network negotiation
   algorithm will have sufficient information to choose on behalf of
   the user agent, thereby optimizing the negotiation process.  A good
   strategy for dynamic extension would be to extend the request with
   with those media types, languages, charsets, and feature tags
   mentioned in the variant lists of past list responses from the
   server.  The Accept header in the extended request, if present,
   will generally contain "*/*" to indicate that only partial
   information is given in this header.  The Accept-Features header,
   if present, will generally contain a "*" for the same reason.


12 User agent support for transparent negotiation

   This section specifies the requirements a user agent must satisfy
   in order to support transparent negotiation.  If the user agent
   contains an internal cache, this cache should satisfy the
   requirements for proxy caches in Section 14.


12.1 Handling of responses

   If a list response is received when a resource is accessed, the
   user agent must be able to automatically choose, retrieve, and
   display the best variant, or display an error message if it is not
   capable of displaying any of the available variants.  The
   negotiation algorithm which chooses the best variant (or chooses to
   display an error message) need not be the network negotiation
   algorithm.

   If a choice response or ad hoc response is received when a resource
   is accessed, the usual action is to automatically display the
   enclosed entity.  A user agent may however choose to apply its
   (custom) negotiation algorithm to the received variant list, and to
   automatically retrieve and display another variant if the algorithm
   indicates that this other variant has a better quality.

   When receiving a choice response, a user agent should check if the
   if the negotiable resource and the chosen variant resource are
   neighbors.  If this is not the case, the user agent should reject
   the choice response as a probable spoofing attempt and display an
   error message, for example by internally replacing the choice
   response with a 502 (bad gateway) response.


12.2 Presentation of a transparently negotiated resource

   If the user agent is displaying a variant which is not an embedded
   or inlined object and which is the result of transparent
   negotiation, the following requirements must be met.

    1. The user agent should allow the user to review a list of all
       variants bound to the negotiable resource, and to manually
       retrieve another variant if desired.  There are two general
       ways of providing such a list.  First, the information in the
       Alternates header of the negotiable resource could be used to
       make an annotated menu of variants.  Second, the entity
       included in a list response of the negotiable resource could be
       displayed.  Note that a list response can always be obtained by
       doing a GET request on the negotiable resource with a Negotiate
       header but without Accept headers.

    2. The user agent should make available though its user interface
       some indication that the resource being displayed is a
       negotiated resource instead of a plain resource.  It should
       also allow the user to examine the variant list included in the
       Alternates header.  Such a notification and review mechanism is
       needed because of privacy considerations, see Section 15.1.

    3. If the user agent shows the URI of the displayed information to
       the user, it should be the negotiable resource URI, not the
       variant URI that is shown.  This encourages third parties, who
       want to refer to the displayed information in their own
       documents, to make a hyperlink to the negotiable resource as a
       whole, rather than to the variant resource which happens to be
       shown.  Such correct linking is vital for the interoperability
       of content across sites.

    4. Similarly, if the user agent stores a reference to the
       displayed information for future use, for example in a hotlist,
       it should store the negotiable resource URI, not the
       variant URI.

   It is encouraged, but not required, that some of the above
   functionality is also made available for inlined or embedded
   objects, and when a variant which was selected manually is being
   displayed.


13 Origin server support for transparent negotiation

   This section covers origin server support from a HTTP viewpoint.
   see Appendix 19.2 for implementation considerations.


13.1 Requirements

   To implement transparent negotiation on a resource, the origin
   server must be able to send a list response when getting a GET
   request on the resource.  It should also be able to send
   appropriate list responses for HEAD requests.  The origin server
   must always send a list response for a request with a Negotiate
   header but without Accept headers.

   The origin server may also return choice responses and ad hoc
   responses, except in the special case above.  The negotiation
   algorithm used need not be the network negotiation algorithm.

   Negotiability is a binary property: a resource is either
   transparently negotiated, or it is not.  Origin servers should not
   vary the negotiability of a resource, or the variant list bound to
   that resource, based on the request headers which are received.
   The variant list and the property of being negotiated may however
   change through time.  The Cache-Control header can be used to
   control the propagation of such time-dependent changes through
   caches.

   It is the responsibility of the author of the negotiable resource
   to ensure that all resources in the variant list serve the intended
   content, and that the variant resources do not engage in
   transparent content negotiation themselves.

   An origin server may shorten a list, choice, or ad-hoc response to
   a 304 (Not Modified) or 206 (Partial Content) response if an
   If-No-Match or If-Range header in the request message allows it
   according to the HTTP/1.1 specification [1].


13.2 Negotiation on transactions other than GET and HEAD

   If a resource is transparently negotiable, this only has an impact
   on the GET and HEAD transactions on the resource.  It is not
   possible (under this specification) to do transparent content
   negotiation on the direct result of a POST request.

   However, a POST request can return an unnegotiated 303 (See Other)
   response which causes the user agent to do a GET request on a
   second resource.  This second resource could then use transparent
   content negotiation to return an appropriate final response.  The
   figure below illustrates this.

      Server ______ proxy ______ proxy ______ user
      x.org         cache        cache        agent

        < -------------------------------------
        |     POST http://x.org/cgi/submit
        |     <form contents in request body>
        |
        -------------------------------------- >
              302 See Other                    |
              Location: http://x.org/result/OK |
                                               |
        < -------------------------------------
        |     GET http://x.org/result/OK.nl
        |      small Accept headers
        |
      able to choose on
      behalf of user agent
        |
         ------------------------------------- >
              choice response with             |
              ..result/OK.nl variant           |
                                           displays OK.nl

   See the HTTP/1.1 specification [1] for details on the 303 (See
   Other) response code.  Note that this response code is not
   understood by most HTTP/1.0 clients.


14 Proxy support for transparent negotiation

   Transparent content negotiation is designed to work through proxies
   which only implement the HTTP/1.1 specification [1].  Thus, in a
   sense, all HTTP/1.1 proxies support transparent content
   negotiation.

   Plain HTTP/1.1 allows proxies to cache list, choice, and ad hoc
   responses, and to efficiently revalidate them by using the
   If-None-Match header.  This specification defines additional
   optimization mechanisms.

   First, a proxy may sometimes reuse an Alternates header taken from
   a previous response (Section 10.3) to negotiate on behalf of the
   user agent or origin server and to construct choice responses
   (Section 10.2).  Proxies must not use custom negotiation
   algorithms, they may only use the network negotiation algorithm.

   Second, if a proxy receives a choice response, it may extract and
   cache the normal response embedded therein, as described in
   Section 10.4.

   Third, a proxy may shorten a list, choice, or ad-hoc response to a
   304 (Not Modified) or 206 (Partial Content) response if an
   If-No-Match or If-Range header in the request message allows it
   according to the HTTP/1.1 specification [1].


15 Security and privacy considerations

15.1 Accept headers revealing information of a private nature

   Accept headers, in particular Accept-Language headers, may reveal
   information which the user would rather keep private unless it will
   directly improve the quality of service.  For example, a user may
   not want to send language preferences to sites which do not offer
   multi-lingual content.  The transparent content negotiation
   mechanism allows user agents to omit sending of the Accept-Language
   header by default, without this affecting the outcome of the
   negotiation process if transparently negotiated multi-lingual
   content is accessed.

   However, even if Accept headers are never sent, the automatic
   selection and retrieval of a variant by a user agent will reveal a
   preference for this variant to the server.  A malicious service
   author could provide a page with `fake' negotiability on
   (ethnicity-correlated) languages, with all variants actually being
   the same English document, as a means of obtaining
   privacy-sensitive information.  Such a plot would however be
   visible to an alert victim if the list of available variants and
   their properties is reviewed.

   Some additional privacy considerations connected to Accept headers
   are discussed in [1].


15.2 Spoofing of responses from variant resources

   The caching optimization in Section 10.4 gives the implementer of a
   negotiable resource control over the responses cached for all of
   its variant resources which are neighbors.  This is a security
   problem if a neighboring variant resource belongs to another
   author.  To provide security in this case, the HTTP server will
   have to filter the Content-Location headers in the choice responses
   generated by the negotiable resource implementation.


16 Acknowledgments

   Work on HTTP content negotiation has been done since at least 1993.
   The author is unable to trace the origin of many of the ideas
   incorporated in this document.  This document builds on an earlier
   specification of content negotiation as recorded in [2].  Many
   members of the HTTP working group have contributed to the work
   reflected in this document.  The authors wish to thank the
   individuals who have commented on earlier versions of this
   document, including Brian Behlendorf, Daniel DuBois, Ted Hardie,
   Larry Masinter, and Roy T. Fielding.


17 References

   [1] Fielding et al, Hypertext Transfer Protocol -- HTTP/1.1.
       Internet-Draft draft-ietf-http-v11-spec-06.txt, HTTP Working
       Group, July 4, 1996.

   [2] Roy T. Fielding, Henrik Frystyk Nielsen, and Tim Berners-Lee.
       Hypertext Transfer Protocol -- HTTP/1.1.  Internet-Draft
       draft-ietf-http-v11-spec-01.txt, HTTP Working Group, January,
       1996.

   [3] T. Berners-Lee, R. Fielding, H. Frystyk, Hypertext Transfer
       Protocol -- HTTP/1.0.  RFC 1945.  MIT/LCS, UC Irvine, May 1996.

   [4] K. Holtman, A. Mutz, Feature Tag Registration Procedures.
       Internet-Draft draft-ietf-http-feature-reg-00.txt, HTTP
       Working Group, October 30, 1996.


18 Authors' addresses

   Koen Holtman
   Technische Universiteit Eindhoven
   Postbus 513
   Kamer HG 6.57
   5600 MB Eindhoven (The Netherlands)
   Email: koen@win.tue.nl

   Andrew H. Mutz
   Hewlett-Packard Company
   1501 Page Mill Road 3U-3
   Palo Alto CA 94304, USA
   Fax +1 415 857 4691
   Email: mutz@hpl.hp.com


19 Appendices

19.1 Adding an Expires header to ensure HTTP/1.0 compatibility

   To ensure compatibility with HTTP/1.0 caching proxies which do not
   recognize the Vary header, an Expires header with a date in the
   past can be be added to the response, for example

        Expires: Thu, 01 Jan 1980 00:00:00 GMT

   If this is done by an origin server, it should usually also include
   a Cache-Control header for the benefit of HTTP/1.1 caches, for
   example

              Cache-Control: max-age=604800

   which overrides the freshness lifetime of zero seconds specified by
   the included Expires header.


19.2 Origin server implementation considerations

19.2.1 Implementation with a CGI script

   Transparent content negotiation has been designed to allow a broad
   range of implementation options at the origin server side.  A very
   minimal implementation can be done using the CGI interface.  The
   CGI script below is an example.

      #!/bin/sh

      echo $HTTP_NEGOTIATE | awk '$0~"^(|.*,)[\t ]*trans[\t ]*(|,.*)$" \
      { print "Status: 300 Multiple Choices" }' -

      cat - <<'blex'
      Alternates: {"stats.tables.html" 1.0 {type text/html} {features
      tables}}, {"stats.html" 0.8 {type text/html}}, {"stats.ps" 0.95
      {type application/postscript}}
      Vary: *
      Content-Type: text/html

      <title>Multiple Choices for Web Statistics</title>
      <h2>Multiple Choices for Web Statistics:</h2>
      <ul>
      <li><a href=stats.tables.html>Version with HTML tables</a>
      <p>
      <li><a href=stats.html>Version without HTML tables</a>
      <p>
      <li><a href=stats.ps>Postscript version</a>
      </ul>
      blex

   The Alternates header in the above script must be read as a single
   line.  The script generates a list response for user agents capable
   of transparent content negotiation, and an ad hoc 200 (OK) response
   for all non-negotiating agents.

19.2.2 Direct support by HTTP servers

   Sophisticated HTTP servers could make a transparent negotiation
   module available to content authors.  Such a module could
   incorporate implementations of the network negotiation algorithm
   (Section 11) and the algorithm for generating choice responses
   (Section 10.2).  The definition of interfaces to such modules is
   beyond the scope of this document.

19.2.3 Web publishing tools

   Web publishing tools could automatically generate several variants
   of a document (for example the original TeX version, a HTML version
   with tables, a HTML version without tables, and a Postscript
   version), together with an appropriate variant list in the
   interface format of a HTTP server transparent negotiation module.
   This would allow documents to be published as transparently
   negotiable resources.


19.3 Open issues in transparent content negotiation

   At the time of writing of this specification, the following issues
   are still open.

   - The feature tag registration process has yet to be defined.

   - A core set of feature tags (Section 8.2) has not yet been
     defined, though there has been some related work on display
     attributes, see Appendix 19.4.1.

   - Though it is expected that the feature negotiation framework will
     solve many current and future negotiation problems, it is also
     expected that there will remain current and future negotiation
     problems not solved by feature negotiation.  Currently, there is
     little experience in this area.


19.3B Server-side rendering engines

   Cache efficiency considerations generally demand that content
   negotiation is limited to a few variants, and that any fine-tuning
   of a representation (for example the scaling of a title image to
   fit the exact screen width) is done at the user agent side.

   However, some negotiation models may demand the use of server-side
   rendering engines, which fine-tune a representation depending on
   exact parameters supplied by the user agent.  This appendix
   discusses how to extend the transparent negotiation framework with
   a mechanism for reliably delivering such parameters to the server.

   The variant description syntax can be extended with an optional
   sent-tags attribute:

      sent-tags-attribute = "{" "send-tags" 1%ftag "}"

   The presence of this attribute would indicate that the variant URI
   gives access to a server side rendering engine.  An example variant
   description is:

     {"paper.server-render.ps" 1.0 {type application/postscript}
         {send-tags paper-size dpi printcolor print-gamma}}

   If this variant is chosen as the best variant and retrieved in a
   GET request, the user agent would extend the
   paper.server-render.ps URI with those feature tags in the
   send-tags attribute it supports, and their associated preferred
   values if applicable:

    GET paper.server-render.ps?paper-size=A4&dpi=600 HTTP/1.1

   Thus, the exact rendering parameters end up at the server side
   rendering engine.  The URI is extended using the form submission
   syntax for downwards compatibility reasons: this allows the list
   response to contain a HTML form giving access to the engine.

   Server-side rendering engines may by interesting as a fallback
   mechanism when new client-side rendering engines are introduced.
   The following variant list illustrates this:

     {"paper.html" 0.8 {type text/html}},
     {"paper.tex"  1.0 {type application/x-tex}
         {features client-render-tex}},
     {"paper.server-render.ps" 1.0 {type application/postscript}
         {features !client-render-tex}
         {send-tags paper-size dpi printcolor print-gamma}}


19.3C Quality factors in the Accept-Features header

   The feature negotiation framework specified in Section 6 does not
   allow for the introduction of new dimensions like the type and
   language dimension, in which the user can express a relative
   preference for each of the elements by sending quality factors in
   the Accept header.  To support the introduction of such dimensions,
   the Accept-Features header syntax could be extended with quality
   factors.  An example would be the following Accept-Features header:

     Accept-Features: paper=a4;q=1.0, paper=a5;q=0.8, !paper=a3, *

   Examples of overall quality factors computed using this header
   would be:

     0.8 for {"x" 1.0 {features paper=a5}}
     1.0 for {"y" 1.0 {features paper=a4/0.95}}
     0.0 for {"z" 1.0 {features paper=a3}}

   Quality factors could also be useful on value-less feature tags,
   for example

     Accept-Features: tables, larger_2x;q=0.7, *

   would degrade the overall quality of any variant which uses the
   `larger_2x' feature with a factor of 0.7.  The quality factor on
   `larger_2x' would express the relative preference of getting a
   variant which is about twice as large as the smallest variant.
   Examples of overall quality factors computed using this header
   would be:

      0.8 for {"small" 0.8 {features tables}}
      0.7 for {"large" 1.0 {features tables larger_2x/1}}

   In this example, the small variant is chosen: the resource author
   assigned a lower source quality to the small variant, but the
   difference in source quality was not big enough to offset the
   user's dislike of getting a variant which is twice as large.  Note
   that the `/1' causes user agents which do not recognize the
   `larger_2x' tag to select the large variant.

   The proposed semantics for quality factors on feature tags are as
   follows.  First, the Accept-Features headers element

       ftag;q=Q

   assigns the quality factor Q to the feature predicate `ftag' made
   true by it, and the elements

       ftag=V;q=Q  ftag={V};q=Q  ftag<=N;q=Q

   assign the quality value Q to all feature predicates `ftag=X' made
   true by them, but not to the plain predicate `ftag'.  All true
   feature predicates not explicitly assigned a quality factor get the
   default factor 1.

   To account for the quality factors on the feature tags, the overall
   quality of the variant is multiplied with the quality factor
   of every `active feature predicate' in the variant description.
   A predicate is active if it

     1. is made true by the Accept-Features header

     2. does not start with !

     3. and, if appearing in an fpred-bag, does not appear after
        another predicate which also satisfies 1. and 2.

   As an example, with

     Accept-Features: aa, !bb, cc;q=0.7, dd, *

   the variant described as

     {"x" 1.0 {features aa/0.8 !xx !bb [!yy cc dd]}}

   has the active feature predicates `aa' and `cc', while the `!bb'
   and `dd' predicates are not active.

       Note: If a user sends quality factors with feature tags not
       specifically designed to have them, this will in general not
       produce improved negotiation results.  For example, depending
       on the particular variant list, sending `animated_gifs;q=0.8'
       to express a mild dislike of animated gifs can easily fail to
       be effective.  Sending `!animated_gives' could be a bit more
       effective, but only a user agent configuration option which
       completely disables animated gifs can deliver good results in
       all cases.


19.4 Other negotiation specifications

19.4.1 User-Agent Display Attributes

   (Comments based on the internet draft
   draft-mutz-http-attributes-01.txt.)

   User-Agent Display Attributes provide a means for an HTTP client to
   inform a server about its display capabilities.  The draft
   mentioned above defines a number of such attributes as feature
   tags.

19.4.2 HTTP/1.2 Extension Protocol (PEP)

   (Comments based on the internet draft draft-ietf-http-pep-02.txt.)

   PEP is a system for HTTP clients, servers, and proxies to reliably
   reason about custom extensions to HTTP.  Its purpose places PEP one
   meta-level above transparent content negotiation, which is intended
   as a (standardized) extension of HTTP.

   Transparent content negotiation and HTTP/1.1 were both carefully
   designed to work together without requiring meta-negotiation.
   Therefore, transparent content negotiation does not require or rely
   on the use of PEP.

   Though PEP and transparent content negotiation both offer
   extensible negotiation frameworks, they operate in different areas
   of negotiation.  PEP negotiates on the use of mutually understood
   protocol extensions, while transparent content negotiation is a
   method for efficiently retrieving the best representation of some
   information.  Thus, deployment and use of PEP can be orthogonal to
   deployment and use of transparent content negotiation.

   Though some elements of PEP resemble some elements of transparent
   content negotiation, it does not seem advantageous to attempt to
   generalize (parts of) both specifications in order to allow sharing
   of code by implementations.  The two specifications seem to have
   more unrelated parts than similar parts.


19.5 Related issues

19.5.1 Current negotiation practice

   In current practice, most user agents send short Accept headers in
   every request.  These headers usually inadequately describe the
   user agent capabilities and user preferences, except maybe for
   inline image requests.  Servers providing multiple representations
   of the same information usually do so under different URIs, and
   allow users to manually select a representation by clicking a
   particular link.

   Some HTTP servers implement extensive negotiation capabilities
   based on HTTP/1.0 headers.  However, with current user agents,
   these capabilities can seldom be used effectively.

   Some origin servers use the contents of the User-Agent request
   header for negotiation purposes.  These contents sometimes allow
   the server to infer information about capabilities which cannot be
   expressed with the existing Accept headers.

   Some sites use Netscape cookies (persistent client state http
   cookies) to implement a user preferences mechanism.

19.5.2 Bandwidth negotiation

   With a bandwidth negotiation mechanism, the time needed to retrieve
   a particular variant over the network can be taken into account
   during the negotiation process.  Work on mechanisms for bandwidth
   negotiation has been done since at least 1993, but this work has
   not yet yielded a successful standard mechanism for bandwidth
   negotiation.  Cascaded proxy caches introduce additional complexity
   in this area.

   This specification does not attempt to solve the problem of
   bandwidth negotiation.  There is some hope that the feature
   negotiation framework will allow the introduction of an adequate
   solution for bandwidth negotiation.


Expires: May 21, 1997