May 18th (1998)

Submitted by Jamie Zawinski to Miscellaneous.

“Intertwingularity is not generally acknowledged – people keep pretending they can make things deeply hierarchical, categorizable and sequential when they can't. Everything is deeply intertwingled.” – Ted Nelson

In the following, I outline a potential project to make it easier to deal with a massive volume of personal messages: excavating, traversing, relating, reporting, annotating.

I call this hypothetical program “Intertwingle.”

• introduction.
• links are legion.
• searches are intersections.
• implementation.
  • parser.
  • database.
  • query tool.
  • presentation tools.
• future.

Intertwingle can be seen as a unification of a search tool and an address book. It is not, however, a mail reader. The presentation of query results could be done through a mail reader, but the intention is that ones choice of mail reader should be orthogonal to the use of this tool. The two kinds of tools just happen to operate on the same data.

The design philosophy is that any time there is a visual representation of an object, the corresponding object should be accessible with a gesture: That chasing links is easier than composing search terms (but both are needed.)

The target audience is individuals who have a lot of mail. The target audience is not inhabitants of the corporation, it is people. Needs which are specific to IS Managers, or to Enterprise Directory Services are not of interest. This is about the general problem of handling lots of individual mail. (Whether in the context of personal mail, or job-related mail, the problem is the same: you've got a lot of it; now what do you do?)

Sharing is an interesting problem, and may be addressed, but I feel it is explicitly secondary in priority to solving the problem in the non-shared domain. (But, we should think about it up front, because that kind of thing tends to be hard to retrofit.) links are legion. The sheer multitude of representations-of-objects yields a colossal number of potential links to follow, which is why I anticipate link-chasing to be a (usually) far easier method of excavation than searching. For example, here are the headers of a typical message:

Date:  Sun, 3 Jul 94 16:40:07 PDT
From:  Jamie Zawinski <jwz@mcom.com>
To:  eng
Subject:  printing
In-Reply-To:  Chris Houck's message of Sun 3-Jul-94 13:19:23 -0700 <9407032019.AA18853@neon.mcom.com>
Message-ID:  <19940703093034.jwz@islay.mcom.com>
References:  <9407032019.AA18853@neon.mcom.com>

There is a great deal of structure there:

Sun, 3 Jul 94 16:40:07 PDT This is a representation of a point in time. From here one can envision traversing to a list of other messages within some range of that moment: that hour, that day, that month, that year.

Jamie Zawinski jwz@mcom.com This is a description of a particular person. From here one should be able to easily get to information related to that person: an address book entry, or a list of all messages sent by them, or sent to them, or any number of other annotations.

Jamie Zawinski This is a name, not a person, and names are notoriously non-unique. From here it would be useful to get to a list of all known people who have claimed that name (from the set of people who are message senders or recipients.)

jwz@mcom.com This is an email address, not a person, and while one email address is usually not used by more than one person, it's quite common for one person to have many email addresses (or many variations on the same address.) From here it would be useful to get to a list of all known people who have used that address (from the set of people who are message senders or recipients) and from there to the set of other addresses used by that person or those people. One might also find it useful to get a list of messages associated with this address (while excluding messages from other addresses of the same person.)

eng This is an email address, yet it happens to be a mailing list. There is no one person associated with it, yet the set of operations one might like to perform on it is very similar.

printing This is unstructured text, and what one does with unstructured text is attempt to match patterns in it. There are any number of other properties associated with this particular piece of text: it is in a header field called Subject in a message from Jamie Zawinski, on Sunday, July 3rd, and so on. All of these are interesting properties that are within one or two link-hops of the text itself. Their proximity is what makes them interesting.

Chris Houck A name, as above.

Chris Houck's message An ambiguous reference to a message. From here, one should be able to get to the set of all messages from someone who claimed the name Chris Houck.

Chris Houck's message of Sun 3-Jul-94 13:19:23 -0700 Another reference to a message, probably less ambiguous.

9407032019.AA18853@neon.mcom.com 19940703093034.jwz@islay.mcom.com These also are references to particular messages, the least ambiguous representations so far; however, they are still slightly ambiguous, since message IDs refer to original messages: there could be multiple copies of these messages with slightly different headers or other annotations within the message-store.

Any any time there is a link, one can imagine an equal but opposite counter-link: when we talk of reaching lists of objects above, the object by which we reached that list will always be a member of the list. And if A is three hops away from D, then D is three hops away from A, and traversal in both directions should be possible.

However, the object at the other end of the link does not necessarily encode the reverse path in its usual visual representation. For example, while messages point to the message to which they are a reply, the parent doesn't (in itself) point to the children. This implicit relationship must be made explicit: it must be easy to get from a message to the set of messages which refer to it. All links must be bidirectional.

Further structure exists outside of the message headers themselves:

• Messages live in folders.
  • Folders have names.
  • Folders are sometimes arranged in a hierarchy.
  • Folders tend to store messages linearly, in a particular order: thus, each message has “previous” and “next” relationships with other messages.
• Messages can contain other messages (forwarded messages, or digests.) Each such message is a message in its own right, but the containment relationship can be important.
• Messages have bodies.
  • The bodies can contain unstructured text.
  • The bodies can contain text that is named, for example, an attached text file which has a file name or description specified in its attachment headers.
  • The bodies can contain binary objects which, while not textually searchable, are named and described.
  • Bodies can contain hyperlinks. Plain-text messages might happen to have detectable URLs in them, and HTML messages have many mechanisms for referring to other objects. This implies that it would be interesting to traverse from a message, to information about a web page that it refers to, and back to a set of messages which refer to objects on that server.

Following a link only gives you one dimension of mobility. A search can be seen as following multiple links, and finding the intersection (or union) of the results of those links.

Any link-relationship should be searchable. For example:

• All messages from person between date and date that have pattern in the body.
• All messages from person which contain a message from person.
• All messages to mailing-list which refer to URL.
• All messages containing text in the main body, but not in an attachment.
• All messages with an attachment whose file name contains string.

The basic components of this system are:

The module which reads the existing message store (directories of BSD mbox files, or news spool directories, or whatever) and parses them into tagged, indexable data.

It needs to understand where messages begin and end, understand how to descend into MIME structures, how to translate HTML into indexable text, how to recognise URLs, and so on, and so on.

It will presumably generate an intermediate data representation which can be more easily fed to the database. A pretty-printed version of the representation of a message might look like this (if you will excuse my lisp-centric upbringing; here in the modern world, this would presumably be done with XML):

(:message
  (:db-id "globally-unique-identifier")
  (:header-field (:key "from")
		 (:addr "Jamie Zawinski" "jwz"))
  (:header-field (:key "newsgroups")
		 (:news "mcom.test"))
  (:header-field (:key "subject")
		 (:text "hey"))
  (:link "http://url-found-in-some-textual-header/")
(:attachment
 (:type "text/plain")
 (:body "message body text")
 (:link "http://url-found-in-body-text")
 (:addr-or-id "email@address.found.in.body")
 (:addr-or-id "or@maybe.its.really.a.message.id"))  
(:attachment
 (:type "text/plain")
 (:name "filename")
 (:description "description")
 (:text "decoded/stripped text of attachment")
 (:link "http://ijkl")
 (:link "http://mnop"))
(:attachment
 (:type "application/postscript"))
(:attachment
 (:type "message/rfc822")
 (:message-pointer "db-id")))

These objects are shallow: that last “db-id” mentioned in the example is a pointer to a top-level message object that will be coming up soon (probably next in the stream.) That is, deeply nested trees of messages are flattened. (An interesting search term might be “depth > 1” for when you're looking for something, and you know it was in a forwarded message, but you don't remember from whom.)

Deeply nested MIME structures (multipart/ forms) are also flattened. Content-Disposition is always assumed to be inline for purposes of indexing; we index the body of any part that is of a text type. There is no special handling for multipart/alternative forms: each part is indexed as for multipart/mixed.

A more formal representation might be

    msg_desc      =  db_id *msg_header
		    *link_part *addr_id_part
		    *msg_body
    msg_header    =  header_name header_body
    msg_body      =  text_part / link_part /
                       attach_part
    header_name   =  keyword
    header_body   =  text / *mailbox /
		    *newsgroup / *msg_id / date
    mailbox       =  name address
    name          =  keyword
    address       =  keyword
    newsgroup     =  keyword
    msg_id        =  keyword
    date          =  
    text_part     =  content_type text
    content_type  =  keyword
    link_part     =  url
    addr_id_part  =  address / msg_id
    url           =  text
    attach_part   =  content_type
		       [attach_name]
                       [attach_desc]
                       [attach_value]
		    *link_part *addr_id_part
    attach_name   =  text
    attach_desc   =  text
    attach_value  =  text_part / db_id
    db_id         =  
    keyword       =  
    text          =  

(Note: I've actually already written this parser; it's not a lot of code, but it seems to work fairly well. If anyone is seriously interested in taking this project and running with it, I'll see about getting permission to release that code.)

The module which stores the output of the parser on disk in some quickly-retrievable format. It needs to have both relational and full-text-indexing properties; many of the searches we want to do could be accomplished with a database that was nothing but a glorified set of hash tables; but body searches need to be done in some more clever way. (Perhaps simply putting every word in a hash table would be sufficient, but I doubt it.) And more to the point, the text searches have to take advantage of the tagging of the data, so that, for example, constraining a search to be in the subject and not the body actually makes the search go faster instead of slower.

Incremental updates are probably pretty important. I doubt we could get away with a setup that required a nightly update.

It seems clear that RDF would be the way go go here.

All of the web search engines force the user to type in boolean expressions. Sometimes that's ok, but we should do something better, that lets the user construct expressions with a GUI.

Drawing on the notion that searches are really set operations, perhaps one aspect of the search tool could be drag-and-drop: to add a set of messages to the union of messages returned, drop the link on the “Or” box. To add it to the intersection of messages returned, drop it on the “And” box. Of course, that doesn't handle deeper boolean expressions, or textual searches. Maybe it's a dumb idea.

There are objects, sets of objects, and presentation tools. There is a presentation tool for each kind of object; and one for each kind of object set.

The presentation tools for these kinds of objects needn't be complicated, since there's not a lot of information to show: just a bunch of links and/or commands. For example, there needs to be a place to hang the “show me all people with this name” gesture, and the “show me all messages from this user” gesture. But just including the list there isn't going to work, since it's long; really, there wants to be a way to initialize a search with this user. Perhaps activating one of those controls would bring up the search tool with some terms already filled in, like

user = “Jamie Zawinski jwz@mozilla.org”

Getting back to the drag-and-drop idea, dragging that button onto an existing search tool could expand the search to include that term.

One should be able to store annotations on people: even something as simple as a single text field would add a great deal of power. These annotations should themselves be searchable. These annotations should be able to contain (clickable!) references to other people or messages or newsgroups or…

BBDB convinces me that this is an absolute requirement.

The problem with the annotation notion is that it's the first time that we consider a piece of data which is not merely a projection of data already present in the message store: it is out-of-band data that needs to be stored somewhere. In the address book? In LDAP? I have no idea.

Perhaps a simple list is sufficient, with options to sort in various ways (by last name, first name, email, host-name, or host-domain.)

Presenting a single message is straightforward: just return a message/rfc822 or text/html document. However, there should be some other controls available: Reply-To-Sender, Reply-To-All, Forward. And there needs to be a place to hang the reciprocal links to the referring messages, to the folder, and so on.

Annotations of messages would be interesting as well. For example, one might want to make a note to one's self that two messages from different people refer to the same issue and should be dealt with at the same time.

This presentation has to be fairly powerful; it needs to present a decent summary of the messages (with resizable columns for sender, recipient, date, and so on) and be able to do all the usual sorting and threading tricks. Basically, this has to be a very good thread display.

It should also be able to incrementally update as results are coming back from the database, so that the user can see the results they're getting (and even examine messages) while more results are still coming in.

Note that, to this view, the concept of “folder” is meaningless: a folder name is just another property by which searches can be pruned.

Today, I can point my “message set browser” at my Inbox folder, but I can't point it at the set of messages with word in the body. The special treatment of Inbox is arbitrary and limiting.

Annotating a message-set could mean manually including and excluding specific messages: a message-set could be considered a “bucket” which the user can then manipulate by hand, assign a name, and keep around. For use as a “to do” list, say. (Message inclusion and exclusion could be handled by manipulating the search terms, so it's not as hard a problem as textual annotations in general.)

Presentation tools should be linked as well: one should be able to pick up the sets displayed in one tool and project them into another. For example:

• Show me all messages with word in body.
• Drag the sender column away: that's a set of people, therefore it is displayed using a “people browser”.
• In the people browser, click on an address: refine the search to contain only those in the same domain as that address. A new, smaller list of people is presented.
• Project the addresses of those people into a message-set-viewer: this shows all mail received from any of those people.

Perhaps the message-set presentation is a simulated IMAP folder. Perhaps the message and message-set presentation tools are a mail reader.

However, every element of the display needs to be deeply intertwingled with the database. Simply dropping the messages into a mail reader would defeat the purpose, which is that every structured piece of text on the screen should be a hyperlink.

The presentation tools could be implemented as client-side Java, or partly as client-side Java and partly as server-generated HTML. (It seems unlikely that the message-set presentation could be implemented solely in HTML, though that's conceivable for the other presentations.)

The other components are server-side, and need to be far faster than Java is currently capable of. And hopefully we won't have to write the database at all, but can just use something publicly available.

There are other interesting data-visualization possibilities here as well; since really what we have is nodes and connections between them, tools like graphers and histogram charts might be applicable as well, to answer questions like

• show me a graph of the age-distribution of my unanswered mail, or,
• show me a graph of people who are known to have directly exchanged mail with each other so that I can see the “clumping” of my correspondents.

The object/presentation infrastructure should be designed so that new tools drop in easily, with few interdependencies.

This sort of model is not applicable merely to the domain of messages; it applies equally well to any corpus which has structured, potentially-ambiguous references (or rather, representations of references.)

For example, source code.

Copyright © 1998-2003 The Mozilla Organization. Last modified November 10, 1998