FramerD Concepts
Ken Haase

Quick ref: [OIDs and the Object Database]  [Persistent Indices]  [Frames in FramerD]  [Searching for Frames]  [The DTYPE Protocol

Architecture FramerD has a layered architecture
OIDs and the object database FramerD provides a simple object database which associates numeric object identifers (OIDs) with DType structures. DType structures can contain these object identifiers so that one DType structure can point to the location in the object database where another DType structure is stored.

One of the major components of FramerD is a distributed persistent object database. Objects in the database are designated by unique object identifiers or OIDs. The allocation of OIDs is managed so that different users, projects, or programs will not use the same OID for different purposes.

An OID is a 64-bit address pointing into a virtual memory space whose locations are associated with complex structured objects. These objects can also contain pointers to other OIDs. The idea is that one program can associate an OID with a particular value (for instance, the integer 3 or the list ("Three" THREE 3)). Other programs can retrieve this value (or whatever value replaces it) by just knowing the OID to which it was assigned. OIDs can be included in other structures, written to files, and accessed through indices, so that application users need not typically worry about the exact numerical value of an OID.

OID's numeric IDs are further organized into continuous ranges called pools. Pools are used to organize the storage and management of the values associated with OIDs. There are currently two kinds of pools: file pools are disk files containing DTYPE representations for the values associated with a particular range of OIDs; network pools are network servers using the DTYPE protocol which likewise provide access to a particular range of OIDs.

OID values

OIDs are associated with values which applications may access or modify. When an application makes changes to an OID's value, those changes are local to the application until they are committed, at which point they are saved to the appropriate file or network pool. When this occurs, the new value will be visible to other applications which request it.

Access and revision control is managed by locking particular OIDs. When an application locks an OID, two things are guaranteed: that the value which the application sees for the OID is the latest value; and that the value will not be changed by any other application until either the OID is committed or it is unlocked. Locking of OIDs is normally invisible to applications. When an application sets an OID's value locally, it locks it. When the value is committed, it unlocks it.

For file pools, locking any OID in the pool locks the entire pool; for network pools, OIDs are locked on an individual basis.

Pools reflect both implementational and administrative properties: the values associated with the OIDs in a particular pool are generally provided by a particular file or server; the allocation of new OIDs and the modification of the value for existing OIDs is generally the responsibility of one administrative entity.

External Representations

OIDs have two different external representations. A literal OID reference has the form @hi/lo where hi is the high 32 bits of the address and lo is the low 32 bits, both in hex. For example, an OID that looks like this @1/1b54e refers to the address 0x000000010001b54e. A logical OID reference has the form @/pool-id/offset and indicates an OID in a particular pool. Thus, the reference @/brico/1b54e refers to the same object shown above if the pool brico starts at @1/0. In addition, when the system can figure out a name for an OID, it appears after the @-expression identifying its address, e.g.:

@/brico/1b54e"NOUN.COGNITION synset for example, illustration, instance, and representative"

to make it less cryptic to the user.

The high 32 bits of an object identifer indicate the super pool of the OID. The OID above, for instance, is in the super pool 0x1. The allocation of new object identifiers is based on the division of the 64 bit address space into smaller pools. The first such division is into super pools and subsequent divisions divide super pools into smaller chunks. Each chunk consists of 2i object identifiers.

Getting Pools, OIDs, and their values

The Lisp programmer knows the value of everything but the cost of nothing.
Alan Perlis

The FDScript procedure use-pool arranges for the current session to access a particular pool of OIDs. It's single argument is either a filename or a network specification of the sort used for remote evaluation (e.g. `'). Once use-pool has been called, the value of an OID in the pool can be extracted by the procedure OID-VALUE e.g.

[fdscript] (USE-POOL "")
[11:57:27 Session id=fdscript /OS:DIGITAL UNIX /Compiled:Feb 10 1997 /Started:Tue Feb 11 11:57:27 1997]
[11:57:27 Added pool]
[#POOL @1/0+1048576 {}]
[fdscript] (OID-VALUE @1/33334)
[SPELLING: "cause_to_tip" OBJ-NAME: "cause_to_tip" 
 SENSES: @/brico/32f96"VERB.MOTION synset for tip, cause_to_tip, and cause_to_tilt"
  (V 1 @/brico/32f96"VERB.MOTION synset for tip, cause_to_tip, and cause_to_tilt")]

One can use the procedure random-oid to pick a random OID from a pool, e.g.

[fdscript] (RANDOM-OID (USE-POOL ""))
@/brico/58e1"ADJ.ALL synset for functional"
[fdscript] (RANDOM-OID (USE-POOL ""))

Pools can be stored in variables and used as arguments, e.g.

[fdscript] (define WORDNET (USE-POOL ""))
[fdscript] (RANDOM-OID wordnet)
@/brico/148d3"NOUN.ARTIFACT synset for stoup and stoop"

Also, procedures which expect pools as arguments (like ALLOCATE-OID) will also take strings and automatically interpret them as pool specifications), for example:

[fdscript] (RANDOM-OID "")
@/brico/1d9a7"NOUN.COMMUNICATION synset for hillbilly_music"

Allocating and changing OIDs

The procedure ALLOCATE-OID returns an OID from a pool; pools can be either stored in files on disk or servers on a network. So we can say:

[fdscript] (ALLOCATE-OID (USE-POOL ""))
[11:59:57 Added pool]

to allocate an OID in the network pool "". (This pool is maintained for demo purposes at the Media Laboratory. Values stored in this pool are subject to erasure whenever our disk space gets tight). If we call ALLOCATE-OID again, we get a different oid:

[fdscript] (ALLOCATE-OID (USE-POOL ""))

When an OID is newly allocated it has no values assigned to it, E.G. a call to OID-VALUE

[fdscript] (OID-VALUE @/temp/40001)

fails by returning no results. One can set the value of an OID with the SET-OID-VALUE! procedure:

(SET-OID-VALUE! @/temp/40001 "My first OID")

allowing us to get the value we put there:

[fdscript] (OID-VALUE @a/40001)
"My first OID"

Saving and Reverting OID values

When you modify the value associated with an OID, the value is only changed locally until the changes are committed to the pool containing the object. There are three basic `commitment' procedures for OIDs and pools:

saves the changes associated with a particular OID.
saves all the changes to a particular pool;
saves all the changes to OIDs in the current session, and

All changes are automatically committed if your FDScript session exits normally (and lost if it exits abnormally).

Not all OIDs can be changed or modified. If you change an OID in a file pool, the system attempts to lock the corresponding file before making the change. The file stays locked until your process exits or the file is explicitly unlocked. If you cannot lock the file --- because someone else has locked it or the file is read-only --- you cannot change the value associated with the OID stored in the file.

If you change an OID in a network pool, the system attempts to lock the individual OID before making the change. If this succeeds, the local value for the OID is changed and the remote OID will be unlocked when your changes are committed. If this fails, (for instance, the network server will not permit you to lock the OID), the OID cannot be modified. A server might refuse to lock an OID because another user has locked the OID or because the OID is declared read-only.

Before you've committed your changes, it is possible to back out of the changes by reverting the modifications:

undoes your changes to the individual OID oid.
undoes all the changes to OIDs in a particular pool
undoes all the changes to OIDs in the current session that have not yet been committed.

Two caveats: remember that committed changes cannot be reverted in this way; also, only changes that directly affect the OID or its value can be reverted. If you store a pointer to an OID in a hashtable, for instance, reverting the OID will not remove the hash table entry.

Making file pools

The frames created above were allocated and stored on a demonstration server at MIT. Though the values might stay around for a few days, they are unlikely to last longer for administrative reasons. To create objects with more presistence, one has to either use file pools maintained in files on the local file system or use network pools maintained by oneself or others with more persistence. Actually, the network servers you would use will rely on pools maintained in files on their local file system, so eventually someone has to worry about how OIDs live on disk in file pools. Fortunately, dealing with file pools is straightforward.

The procedure MAKE-FILE-POOL creates a file pool in the local file system, E.G.

[fdscript] (MAKE-FILE-POOL "test.pool" 32)
[#POOL test.pool @a/140000+0/32 {}]

creates a file pool containing 32 possible and no actual objects and and whose first object will be @a/140000. We can allocate that first object with allocate-oid:

[fdscript] (allocate-oid "test.pool")
@/test/0 ; or @a/140000 if printed "literally"

Creating a pool automatically causes the pool to be used, just as if you had called use-pool on the argument.

The command-line program make-file-pool also creates file pools; it looks just like the FDScript call, but without the parentheses, e.g.

sh% make-file-pool test.pool 32
Created test.pool with space for 32 OIDs starting at @a/100

This command line program is written in and invokes FDScript, but it doesn't require that the user deal with the listener and associated parentheses. It also generates a message describing what it has done.

Frames in FramerD Frames are at the heart of FramerD, providing a way to build complex extensible descriptions and to share pointers to these descriptions between different databases and applications.

The word "frame" in FramerD comes from its support for frames, a popular AI representation model. A frame in FramerD is an OID whose value consists of a set of slots. Each slot consists of a slotid --- either a symbol or an OID --- and a set of values. Slots divide the description of objects into several labelled parts. For instance, the following frame from the FramerD port of WordNet represents the concept (synset) of an "example":

SYNSET-ID:  3946824 [[[Slot values may be numbers or strings.]]]
WORDS: (4 values) [[[A slot may have any number of values]]]
DESCRIPTION:  "a single item that is representative of a type"
OBJ-NAME:  "NOUN.COGNITION synset for example, illustration, instance, and representative"
    HYPERNYM:  @/brico/b709"NOUN.COGNITION synset for information"
    HYPONYM: (5 values) [[[The values can also be pointers to other frames]]]
       @/brico/1b56e"NOUN.COGNITION synset for exception"
       @/brico/1b56f"NOUN.COGNITION synset for precedent and case_in_point"
       @/brico/1b570"NOUN.COGNITION synset for quintessence"
       @/brico/1b571"NOUN.COGNITION synset for sample"
       @/brico/1b572"NOUN.COGNITION synset for specimen"
    A:  @/brico/b709"NOUN.COGNITION synset for information"

The description above lists a number of slots, each of which associates a particular slotid (e.g. HYPERNYM) with a set of slot values (which are abitrary objects). This set of associations is called a slotmap. A frame is an OID whose value in the object database is a slotmap. Using OIDs allows different programs and databases to keep separate references to a frame while managing to share changes and augmentations.

Operations on slots include getting their value(s), adding a value, removing a value, and checking whether they contain a particular value. The value may be non-deterministic (a choice), in which case we may refer to the values of the slot.

When the slotid is a symbol, operations on the value of the slot are just operations on the slot's data. Getting the slot's value retrieves the slot data and adding a value to a slot adds an element to the slot's data (which is represented as a non-deterministic value). Except for this last point, frames whose slotids are symbols function much as property lists of symbols in languages like LISP or associative arrays in Perl. (From a computational standpoint, slotmaps are not a very efficient way to store large numbers of associations. For that, it is better to use a hashtable or an external index.)

When the slotid is itself a frame (i.e. an OID whose value is a slotmap), operations on the slot are more complicated. Instead of just operating on the slot's data, operations evaluate methods for computing or testing for values, adding or removing elements, or checking that the given frames or values are correct. The methods are expressions in FDScript, the FramerD scripting language, which can access the variables frame, slotid, data, and value which contain the details of the particular slot and (when appropriate) the value being added, removed, or tested for. The expressions are stored in the following slots of the slotid:

Methods can operate on slots of the same or other frames and normally this simply invokes the corresponding methods, except when this is likely to recur infinitely. This can occur when the inference relations between slots are implicitly circular. For example, slots describing the width, height, and area of a rectangle might be defined in terms of one another. However, this could lead to an infinite recursion if, for instance, by the following path:

To avoid this recursion, operations on a complex slots always check whether an identical operation (getting the same slot, adding, removing, or testing for the same value) is already being performed. If it is already being performed, the recursive operation either:

This allows slots to freely refer to one another without concern for infinite recursion.

Getting at Frames from FDScript

The procedure FRAME-CREATE creates a frame and requires specifying a pool in which the frame's OID will be allocated. The value returned by FRAME-CREATE is this OID. If the pool argument is #f (false), a raw slotmap (rather than an OID whose value is a slotmap) is returned. All of the frame procedures will take a raw slotmap as an argument, but the unique identity of the slotmap will not be preserved between FramerD sessions.

This is a subtle but important point. If you save a raw slotmap, subsequent changes to the slotmap will not be shared. If you save a frame (an OID whose value is a slotmap), subsequent changes will be changed. In general, it is important to use frames for descriptions which may persist beyond the current session.

In addition to the pool argument, frame-create takes a number of other arguments. If it is given one argument, it should be a slotmap to which the corresponding OID's value is initialized. If it is given more than one argument, they are intepreted as a series of slots and values initially assigned to the frame, e.g.

[fdscript] (frame-create ""
              'obj-name "test" 'test-slot 'test-value
              'another-test-slot "another-value")
@a/40010"test" ;<- Prints out using the obj-name slot

the slot OBJ-NAME is used by FDScript when displaying the frame's OID, as you can see in the printed value above.

The procedure FGET extracts a slot from a frame given a slotid, E.G.

[fdscript] (fget @a/40010"test" 'test-slot)

When the slotid is a symbol, FGET simply retrieves whatever was initially or subsequently stored in the slot. The rules are different when the slotid is itself a frame, when certain inferences may happen, but we discuss that below.

The procedure FADD! adds a new value to a frame for a given slotid, E.G.

[fdscript] (fadd! @a/40010"test" 'test-slot 'another-value)
[fdscript] (fget @a/40010"test" 'test-slot)

A formatted description of a frame can be gotten with the procedure FDD:

[fdscript] (FDD @a/40010)
The frame @a/40010:
OBJ-NAME:  "test"
TEST-SLOT: (2 values)
ANOTHER-TEST-SLOT:  "another-value"

FDD is also a command line program with the same name, so you can do fdd (lowercase for Unix) at the shell.

Since frames are OIDs, changes to frames are not permanent until either the session exits normally or the changes are explicitly committed. The procedures COMMIT-OID, COMMIT-POOL, and COMMIT-POOLS all work to save changes to frames and the additional procedure COMMIT-FRAME is just another name for COMMIT-OID.

The procedure FRAME-SLOTS returns all the slotids associated with a frame, E.G.

[fdscript] (frame-slots @a/40010)
{obj-name test-slot another-test-slot}

For example, The following code converts a frame into an association list (i.e. a list of key value pairs):

[fdscript] (define (get-alist-entry unit slotid)
             (cons slotid (set->list (frame-get unit slotid))))
[fdscript] (define (frame->alist frame)
             (set->list (get-alist-entry frame (frame-slots frame))))
(frame->alist @a/40010)
((ANOTHER-TEST-SLOT "another-value") 
 (OBJ-NAME "test"))

Inferences over slots

When a slotid is another frame, procedures like fget and fadd! act a little differently based on the slots of that frame. Each operation has a special slot associated with it and that slot contains a set of FDScript expressions which may be evaluated when the operation is performed:

fget GET-METHODS frame, slotid, data
ftest TEST-METHODS frame, slotid, data, value
fadd! ADD-EFFECTS frame, slotid, data, value
fzap! ZAP-EFFECTS frame, slotid, data, value
when these values are evaluated, certain variables are bound:
is bound to the frame being operated upon
is bound to the slot being operated upon
is bound to primitive values associated with the slotid
is bound to the value being tested for, added, or removed

allowing the procedure to compute extra values, make additional changes, or notify users of changes. For instance, the following definition..

[fdscript] (frame-create test-pool 
             'obj-name "n-hyponyms"
             'get-methods (set-size (GET frame 'hyponyms)))

so that we can ask:

[fdscript] (fget @/brico/1b54e @/brico/3338"n-hyponyms")
Persistent Indices Indices are a way of storing and manipulating associations between objects and descriptions. They can be used to go from words to their roots (e.g. "flew" to "fly") or from features to findings (e.g. "BROTHER-OF Ken" to "Bruce"). FramerD's indices are persistent incremental data structures designed to support millions of keys.

Indices are persistent

The mappings described by an index are maintained outside of the application using the index. An application can change an index and those changes will be automatically available to later instantiations of the application or even of other applications. Changes are not visible at once, however; an application must explicitly commit its changes to make them persistent and shared.

Because indices are persistent, they are typically used as resources by programs and may represent the accumulation of results from many computations and many different processes. A suite of applications may have a set of indices which they share and update.

Indices are demand-driven

When an application uses an index, it normally only uses resources for a fraction of the index. The rest of the index is maintained --- on a remote server or in the file system --- for when it is needed. When neccessary, it goes to the server or file and retrieves the needed mappings. A quite small application can access a huge index through this method.

FramerD has two sorts of built-in indices. File indices store mappings in a file on a local or remote disk and retrieve mappings by random access in the file. Network indices use a simple protocol to access DType servers which provide mappings to their clients. These two types are interchangable.

Indices are incremental

Indices are designed to efficiently support incremental changes. In particular, it is supposed to be efficient to both add new values to existing keys and to add initial values to new keys. This makes index files useful for storing inverted indices used in many information retrieval applications. It is also easy to get the number of values associated with a particular key, which is important for certain kinds of statistical retrieval algorithms.

Indices also cache values in two directions. When an index goes to a file or server to get a mapping, it caches the result locally so that a subsequent request can proceed much faster. For this reason, some FramerD applications may speed up over time as they cache commonly referenced index keys.

When an index is modified, the modifications are stored locally with the application making the modifications. Only when the index is committed do the changes go into the files on disk or across the network.

Indices can index objects

Indices are the basis of a general object indexing facility. This uses the convention that keys of the form: (slot . value) refer to frames whose slot contains value. For example, the key (year-of-birth . 1961) would be associated with all the frames whose year-of-birth slot was 1961. These indices can be used to find frames with particular properties or combinations of properties as well as for more flexible "fuzzy" searches.

Using indices from FDScript

Indices are accessed by the function USE-INDEX:

[fdscript] (USE-INDEX "/local/test-index")
[#INDEX "/local/test-index"]

which can also take a network server specification, as in:

[fdscript] (USE-INDEX "testi@somehost")
[#INDEX "testi@somehost"]

The functions for accessing an index are just like the functions for accessing a hashtable:

(INDEX-GET index key)
gets the values associated with key in index.
(INDEX-ADD! index key new)
adds new to the values associated with key in index.
(INDEX-SET! index key new)
makes new be the only value associated with key in index
(INDEX-ZAP! index key)
ensures that no values will be associated with key in index.
(INDEX-KEYS index)
returns all of the keys stored in an index.

Changes to indices are like changes to pools in that they need to be committed to be permanent. Changes are committed automatically if the FDScript session exits normally. Changes can also be committed manually:

saves the changes made all the keys in index
saves the changes made to all the keys in all of the indices

As with pools, changes to indices can be reverted:

removes all the changes made to keys in index.
removes all the changes made to keys in all indices.

As with pools, changes cannot be reverted once they are committed and reversion applies only to the associations in the index and not to other relations involving the keys or values.

File indices

There are a number of special functions for dealing with file indices.

(make-file-index filename minsize)
Creates a new file index in the file filename that contains at least minsize slots. Note that file indices have a fixed size and need to be explicitly grown using maintenance tools built into FDScript or available from the command prompt. This is because growing a file index with millions of keys can be very time consuming and couldn't be "transparent" even if you wanted it to be.
(cache-index index)
When index is a file index, this keeps a copy of the files offset table in memory, which improves access time by avoiding one disk seek. This is especially useful when disk seeks are particularly slow (for instance with CD-ROM drives).
Arranges for file indices to be automatically cached when referenced. This is a good idea for programs which do a lot of index accesses or for speed when the index file is on a slow device (such as a CD-ROM or network disk).

In addition, there are several command line utilities for dealing with file indices:

make-file-index filename minsize
Creates a file index with (at least) a particular size.
analyze-index filename --keys | --values | --stats
Returns information about the index file in filename. The options can be combined and have the following interpretation:
  • --stats reports hashtable statistics (number of keys, average and max misses per key, most common keys, etc)
  • --keys lists all of the keys with the number of values for each
  • --values lists all of the values
With no arguments, analyze-index simply prints the number of slots in the index.

File indices have standard sizes at roughly powers of 2 and user specified sizes are rounded up to the nearest standard size.

Searching for Frames FramerD builds an object indexing facility on top of the general association facility described above. Indexing a frame stores inverse pointers from its properties to the frame. This allows the frame to be found later based on a set of properties or to find objects similar (i.e. with common properties) to one particular frame.

FramerD builds an object indexing facility on top of the general association facility described above. Indexing a frame stores inverse pointers from its properties to the frame. This allows the frame to be found later based on a set of properties or to find objects similar (i.e. with common properties) to one particular frame.

There are three main ways of indexing frames:

(index-frame index frame)
Indexes all the properties actually stored on frame in index.
(index-frame index frame slot1 slot2... slotn)
Indexes all the specified slots (slot1, slot2, through slotn) on frame. If some of the slots are frames, they may be computed even if they aren't actually stored on frame.
(index-slot-value index frame slot value)
Indexes frame as having SLOT with VALUE (whether it does or not).

Once a frame has been indexed, one can search for it either strictly --- based on certain combinations of slots and values --- or "fuzzily" based on similarity to an instance or set of instances.

Strict Searching

I'll have the french onion soup, without the cheese on top,
the green salad with honey-mustard dressing on the side,
and a decaffeinated coffee with one sugar.

Sally, in When Harry Met Sally

Strict searching uses the FIND-FRAMES procedure:

(find-frames index 
   slot1 value1
   ... slotn valuen)

which finds all objects that have all the specified slots and values. If any of the valuei are non-deterministic sets, the slot need only have one of the specified values, allowing some variations. For example, in the WordNet database, the expression:

[fdscript] (find-frames "" 'words {"hack" "chop"})
{@1/31813"VERB.COMPETITION synset for hack and kick_on_the_arm" 
 @1/151a5"NOUN.ARTIFACT synset for cab, hack, taxi, and taxicab" 
 @1/316ff"VERB.COMPETITION synset for chop and hit_sharply" 
 @1/23e1b"NOUN.PERSON synset for machine_politician, ward-heeler, political_hack, and hack" 
 @1/321a4"VERB.CONTACT synset for chop and strike_sharply" 
 @1/303a5"VERB.CHANGE synset for hack and hack_on" 
 @1/1299a"NOUN.ANIMAL synset for hack, jade, nag, and plug" 
 @1/cdc6"NOUN.ACT synset for chop and chop_shot" 
 @1/31c45"VERB.CONTACT synset for hack and clear" 
 @1/20578"NOUN.PERSON synset for hack, hack_writer, and literary_hack" 
 @1/31c3d"VERB.CONTACT synset for chop and hack" 
 @1/2ee62"VERB.BODY synset for hack and whoop" 
 @1/32cfe"VERB.MOTION synset for chop and move_suddenly" 
 @1/31812"VERB.COMPETITION synset for hack and kick_on_the_shins" 
 @1/129cd"NOUN.ANIMAL synset for hack" 
 @1/1ed62"NOUN.FOOD synset for chop" 
 @1/2262a"NOUN.PERSON synset for hack, drudge, and hacker" 
 @1/31c3f"VERB.CONTACT synset for chop, chop_up, and cut_into_pieces" 
 @1/2f7e3"VERB.CHANGE synset for hack and cut_up" 
 @1/12999"NOUN.ANIMAL synset for hack" 
 @1/bc17"NOUN.ACT synset for chop and chopper"}

finds all the synsets containing either the words "hack" or "chop", while

[fdscript] (find-frames "" 'words "hack" 'words "chop")
@1/31c3d"VERB.CONTACT synset for chop and hack"

finds only the synsets containing both the words "hack" and "chop". One way to think about this visually, is the search performed by find-frames is conjunctive horizontally (along the list of arguments) and disjunctive vertically (within each argument).

Fuzzy Searching

I'll have a burger with fries and a chocolate shake.
Harry in When Harry met Sally

A fuzzy search does not require an exact match, but returns the best possible match measured by the number of overlapping properties. There are a variety of fuzzy search functions as well as a set of tools for writing your own fuzzy search routines. The chief function find-best, looks just like find-frames:

(find-frames index 
   slot1 value1
   ... slotn valuen)

but returns those objects with the largest number of matching properties.

Similarity Searching

I'll have what she's having.
from When Harry Met Sally

One of the most powerful search mechanisms in FramerD and FDscript is "similarity searching" which begins with an object or set of objects and finds objects which have the same properties as those objects, weighting as higher those which are more in common among the set of initial descriptions.

(find-similar indices frame)
Returns the frames indexed in indices which have the most number of slot values in common with frame.
(find-similar indices frameslots)
limits the search for similarity to slots. Returns the frames indexed in indices which have the most number of slot values in common with frame.

For example, the following search finds words with similar meanings to the common sense of "hack" and "chop":

[fdscript] (find-similar ""
             @1/31c3d"VERB.CONTACT synset for chop and hack")
{@1/31be9"VERB.CONTACT synset for shave, trim, and cut_closely" 
 @1/31c3f"VERB.CONTACT synset for chop, chop_up, and cut_into_pieces" 
 @1/31e95"VERB.CONTACT synset for mow and cut_down" 
 @1/325d4"VERB.CONTACT synset for rebate and cut_a_rebate_in" 
 @1/325d7"VERB.CONTACT synset for saw and cut_with_a_saw"}
Fuzzy searching and non-determinism. When the frame argument to find-similar is non-deterministic, the search mechanism does a single search but uses features from each of the frames in combination. As a consequence, the search weighs properties common between the frames more heavily. For example, in this retrieval,
[fdscript] (find-similar "brico@framerd" 
             (amb @1/31c3f"VERB.CONTACT synset for chop, chop_up, and cut_into_pieces" 
                  @1/31e95"VERB.CONTACT synset for mow and cut_down"))
@1/31c3d"VERB.CONTACT synset for chop and hack"

the properties common to the two synsets selects the search pattern which put them together in the first place.

The DTYPE Protocol DTypes are a binary data format and communications protocol underlying FramerD. DTypes allow the storage and transmission of complex recursively structured objects. DTypes can be extended to encode application-specific data while still using generic facilities for storage and search.

DTypes are a portable data representation used throughout the FramerD suite of libraries, tools, and applications. There are two basic levels to the DType representation: a minimal set of "core data types" and an extensible external binary representation for those types and their extensions. Applications and libraries use and extend the native data types, taking advantage of communication, persistence, and indexing facilities which use the external representation.

DTypes differ from other data sharing approaches in not depending on explicit data declarations for sharing complex structures. Distributed processing models like CORBA rely on shared data declarations for communicating complex data structures. DTypes allow a `sloppier' approach where basic data types include nested and labelled heterogenous structures. This allows for fast prototyping of applications and protocols as well as their on-line extension.

The minimal set of native types include fixed and floating point numbers, ASCII strings and symbols, vectors and pair structures. It also includes some special types, especially the OID (Object IDentifier) pointers used by the FramerD object database. The external binary representation is used in three primary ways:

The DType core representation includes the following types:

Native applications can provide their own types and use FramerD storage and indexing facilities by implementing extensions to the external binary representation for their types. These extensions can either be in the form of compounds or package types.

A compound representation consists of a type name (a tag) and a "canonical form" which can be used to identify and regenerate the object. Logically, neither the tag nor the the canonical form can include the object being described or else the translation would recur indefinitely.

A packaged representation consists of a two byte type code followed by 1 or 4 bytes of size information and some amount of additional data. The value of the first byte identifies a "package" for the extension; the value of the second byte (which is defined by the package maintainer) specifies a more precise type code and also specifies the format of the subsequent data in a fashion which FramerD facilities can manipulate without interpreting.

The core FramerD libraries introduce new datatypes with packaged data representations for: