@hviana/sema 0.1.6 → 0.1.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/example/train_base.js +42 -9
- package/dist/src/alu/src/parser.d.ts +9 -0
- package/dist/src/alu/src/parser.js +110 -2
- package/dist/src/canon.d.ts +26 -0
- package/dist/src/canon.js +57 -0
- package/dist/src/geometry.d.ts +13 -2
- package/dist/src/geometry.js +101 -2
- package/dist/src/index.d.ts +1 -0
- package/dist/src/index.js +1 -0
- package/dist/src/mind/attention.js +11 -7
- package/dist/src/mind/learning.js +33 -1
- package/dist/src/mind/match.d.ts +4 -2
- package/dist/src/mind/match.js +30 -6
- package/dist/src/mind/mechanisms/cast.js +18 -4
- package/dist/src/mind/mechanisms/confluence.js +13 -1
- package/dist/src/mind/mechanisms/recall.js +115 -31
- package/dist/src/mind/mind.d.ts +138 -12
- package/dist/src/mind/mind.js +284 -22
- package/dist/src/mind/primitives.d.ts +22 -2
- package/dist/src/mind/primitives.js +95 -6
- package/dist/src/mind/recognition.js +31 -8
- package/dist/src/mind/traverse.d.ts +13 -0
- package/dist/src/mind/traverse.js +41 -0
- package/dist/src/mind/types.d.ts +33 -21
- package/dist/src/store-sqlite.d.ts +10 -0
- package/dist/src/store-sqlite.js +58 -0
- package/dist/src/store.d.ts +23 -0
- package/dist/src/store.js +13 -2
- package/example/train_base.ts +49 -9
- package/index.html +65 -0
- package/package.json +1 -1
- package/src/alu/src/parser.ts +105 -4
- package/src/canon.ts +65 -0
- package/src/geometry.ts +105 -1
- package/src/index.ts +1 -0
- package/src/mind/attention.ts +11 -6
- package/src/mind/learning.ts +39 -1
- package/src/mind/match.ts +29 -6
- package/src/mind/mechanisms/cast.ts +21 -6
- package/src/mind/mechanisms/confluence.ts +15 -1
- package/src/mind/mechanisms/recall.ts +132 -41
- package/src/mind/mind.ts +396 -22
- package/src/mind/primitives.ts +109 -7
- package/src/mind/recognition.ts +36 -8
- package/src/mind/traverse.ts +47 -0
- package/src/mind/types.ts +30 -21
- package/src/store-sqlite.ts +76 -0
- package/src/store.ts +46 -2
- package/test/13-conversation.test.mjs +240 -20
- package/test/35-prefix-edge.test.mjs +86 -0
|
@@ -262,7 +262,21 @@ export async function counterfactualTransfer(ctx, query, pre) {
|
|
|
262
262
|
// corrupted-store read now falls back here instead of voicing a hollow
|
|
263
263
|
// seat into the comparison — the same "empty bytes are no grounding"
|
|
264
264
|
// invariant project() has always enforced.
|
|
265
|
-
|
|
265
|
+
// A node that only ever CONTINUES — an edge SOURCE with no predecessor —
|
|
266
|
+
// is a learnt CONTEXT (a stored question-shaped frame), not an entity.
|
|
267
|
+
// Voicing it verbatim answers a question with a question (the observed
|
|
268
|
+
// cast echo: two stored questions concatenated as an "answer"). The
|
|
269
|
+
// context's role is established by what it LEADS TO, so its seat is its
|
|
270
|
+
// own continuation — the fact — with the reverse projection and the raw
|
|
271
|
+
// bytes as the ordinary fallbacks for genuine entities.
|
|
272
|
+
const seatOfNode = async (id, fallback) => {
|
|
273
|
+
if (ctx.store.prevCount(id) === 0 && ctx.store.hasNext(id)) {
|
|
274
|
+
const fwd = await follow(ctx, id, qv);
|
|
275
|
+
if (fwd !== null)
|
|
276
|
+
return fwd;
|
|
277
|
+
}
|
|
278
|
+
return reverseContext(ctx, id, qv) ?? fallback;
|
|
279
|
+
};
|
|
266
280
|
const seatOf = (p) => seatOfNode(p.anchor, p.ctx);
|
|
267
281
|
const analogs = [];
|
|
268
282
|
for (const p of points) {
|
|
@@ -381,10 +395,10 @@ export async function counterfactualTransfer(ctx, query, pre) {
|
|
|
381
395
|
rNode(ctx, dominant.anchor, "analog", bestSim),
|
|
382
396
|
rNode(ctx, bestAnalog.anchor, "analog", bestSim),
|
|
383
397
|
], [], "the two structures keep distributional company beyond chance — genuine analogs");
|
|
384
|
-
const a = seatOf(dominant);
|
|
398
|
+
const a = await seatOf(dominant);
|
|
385
399
|
const b = bestAnalog.point !== null
|
|
386
|
-
? seatOf(bestAnalog.point)
|
|
387
|
-
: seatOfNode(bestAnalog.anchor, read(ctx, bestAnalog.anchor));
|
|
400
|
+
? await seatOf(bestAnalog.point)
|
|
401
|
+
: await seatOfNode(bestAnalog.anchor, read(ctx, bestAnalog.anchor));
|
|
388
402
|
const answer = await joinWithBridge(ctx, a, b);
|
|
389
403
|
record(answer, "analogical comparison — each analog voiced by the context that establishes its role", new Set([dominant.anchor, bestAnalog.anchor]),
|
|
390
404
|
// A halo-mediated act (the analogy gate) plus two seat projections.
|
|
@@ -70,6 +70,18 @@ export async function confluenceJoin(ctx, query, pre) {
|
|
|
70
70
|
// below.
|
|
71
71
|
const queryWin = pre.queryWindows;
|
|
72
72
|
const queryIds = new Set(queryWin.values());
|
|
73
|
+
// A constraint must bind a CONSTITUENT, not a shard. A genuinely shared
|
|
74
|
+
// form weaves a contiguous RUN of shared discriminative windows — its
|
|
75
|
+
// merged cover span is the form's own length, beyond one perception
|
|
76
|
+
// quantum (2W: the same "two quanta of structure" bar the conjunctive
|
|
77
|
+
// precondition `query.length < 2W` and CAST's weave live under). A long
|
|
78
|
+
// stored document holds thousands of W-windows and will hold a FEW of any
|
|
79
|
+
// query's by accident, but accidental sharing is one window (a merged
|
|
80
|
+
// span of W, at most ~W+overlap bytes: "ow ma", "ys a", "ías " —
|
|
81
|
+
// observed), never a run (" translucent", " featherlight" — the genuine
|
|
82
|
+
// constraints). Shard-bound streams are no constraints, and their meets
|
|
83
|
+
// are connective debris (". Sure,", "ngul" — observed).
|
|
84
|
+
const bindsAConstituent = (cover) => cover.some(([cs, ce]) => ce - cs >= 2 * W);
|
|
73
85
|
const streams = [];
|
|
74
86
|
const rankedCapped = ranked.length > pre.k ? ranked.slice(0, pre.k) : ranked;
|
|
75
87
|
for (const cand of rankedCapped) {
|
|
@@ -96,7 +108,7 @@ export async function confluenceJoin(ctx, query, pre) {
|
|
|
96
108
|
else
|
|
97
109
|
cover.push(curC = [off, off + W]);
|
|
98
110
|
}
|
|
99
|
-
if (cover.length > 0) {
|
|
111
|
+
if (cover.length > 0 && bindsAConstituent(cover)) {
|
|
100
112
|
streams.push({ anchor: cand.anchor, vote: cand.vote, ids, cover, held });
|
|
101
113
|
}
|
|
102
114
|
}
|
|
@@ -6,8 +6,9 @@
|
|
|
6
6
|
import { cosine } from "../../vec.js";
|
|
7
7
|
import { consensusFloor, identityBar, reachThreshold, significanceBar, } from "../../geometry.js";
|
|
8
8
|
import { gistOf, read, resolve } from "../primitives.js";
|
|
9
|
+
import { bytesEqual, indexOf } from "../../bytes.js";
|
|
9
10
|
import { corpusN, hubBound } from "../traverse.js";
|
|
10
|
-
import { project, reverseContext } from "../match.js";
|
|
11
|
+
import { follow, project, reverseContext } from "../match.js";
|
|
11
12
|
import { CONCEPT, STEP } from "../graph-search.js";
|
|
12
13
|
import { unexplainedLabel } from "../rationale.js";
|
|
13
14
|
import { rItem, rNode } from "../trace.js";
|
|
@@ -43,6 +44,44 @@ export async function recallByResonance(ctx, query, pre) {
|
|
|
43
44
|
: "exact self-match — reverse recall to the best-resonating predecessor", nothing, STEP);
|
|
44
45
|
}
|
|
45
46
|
}
|
|
47
|
+
// 0b. ARGUMENT BINDING (RC8): the query is not itself a stored form, but
|
|
48
|
+
// it CONTAINS a recognised constituent that is an edge SOURCE — a learnt
|
|
49
|
+
// pair's left side carried inside a wrapper ("How do you say 'thank you'
|
|
50
|
+
// in French?"). The wrapper is scaffolding; the argument is the span
|
|
51
|
+
// that LEADS somewhere, so its continuation — guided by the whole query's
|
|
52
|
+
// gist — is the answer. Matching the wrapper while ignoring the argument
|
|
53
|
+
// (the observed "good morning" template failure) is worse than silence,
|
|
54
|
+
// so anything short of ONE unambiguous binding falls through: the
|
|
55
|
+
// constituent bar is the same two-quanta (2W) reading confluence binds
|
|
56
|
+
// under, nested recognitions collapse to their MAXIMAL span, and two
|
|
57
|
+
// distinct maximal arguments mean the query asks about neither alone.
|
|
58
|
+
if (qId === null) {
|
|
59
|
+
const W2 = 2 * ctx.space.maxGroup;
|
|
60
|
+
const args = pre.rec.sites.filter((s) => s.end - s.start >= W2 &&
|
|
61
|
+
s.end - s.start < query.length &&
|
|
62
|
+
ctx.store.hasNext(s.payload));
|
|
63
|
+
// Maximal spans by one sorted sweep (starts ascending, ties longest
|
|
64
|
+
// first): every earlier span starts at or before s, so s is contained
|
|
65
|
+
// exactly when the running max end already covers it. O(m log m) — a
|
|
66
|
+
// long input recognises O(|input|) sites, and a pairwise scan here was
|
|
67
|
+
// quadratic in the input.
|
|
68
|
+
args.sort((a, b) => a.start - b.start || b.end - a.end);
|
|
69
|
+
const maximal = [];
|
|
70
|
+
let maxEnd = -1;
|
|
71
|
+
for (const s of args) {
|
|
72
|
+
if (s.end <= maxEnd)
|
|
73
|
+
continue;
|
|
74
|
+
maximal.push(s);
|
|
75
|
+
maxEnd = s.end;
|
|
76
|
+
}
|
|
77
|
+
if (maximal.length === 1) {
|
|
78
|
+
const arg = maximal[0];
|
|
79
|
+
const g = await follow(ctx, arg.payload, queryGist);
|
|
80
|
+
if (g !== null && g.length > 0) {
|
|
81
|
+
return ground(g, "argument binding — the query's sole edge-source constituent, continuation followed", [[arg.start, arg.end]], STEP);
|
|
82
|
+
}
|
|
83
|
+
}
|
|
84
|
+
}
|
|
46
85
|
const whole = await ctx.store.resonate(queryGist, k);
|
|
47
86
|
if (whole.length === 0) {
|
|
48
87
|
return ground(null, "empty store — nothing to resonate with", [], 0);
|
|
@@ -55,9 +94,27 @@ export async function recallByResonance(ctx, query, pre) {
|
|
|
55
94
|
// into bytes — at most one river window (see {@link identityBar}). A
|
|
56
95
|
// fixed cosine bar let long queries claim "near-identical" while whole
|
|
57
96
|
// windows — an answer word — differed.
|
|
58
|
-
|
|
97
|
+
// A hit RESTATES the query when its bytes are the query's own — exactly,
|
|
98
|
+
// or under the response's equivalence (a case/width twin). Restating
|
|
99
|
+
// hits may only conclude through disciplined reverse recall: voicing
|
|
100
|
+
// their bytes echoes the query back at itself (never an answer — the
|
|
101
|
+
// same principle that keeps cast from voicing stored questions), and
|
|
102
|
+
// projecting them forward is reverse recall's containment failure in the
|
|
103
|
+
// other direction — "whatever followed these bytes in some document".
|
|
104
|
+
const qKey = ctx.canon ? ctx.canon(query) : query;
|
|
105
|
+
const restates = (b) => bytesEqual(b, query) ||
|
|
106
|
+
(ctx.canon !== null && bytesEqual(ctx.canon(b), qKey));
|
|
107
|
+
const idBar = identityBar(ctx.store.D, ctx.space.maxGroup, query.length);
|
|
108
|
+
if (top.score >= idBar) {
|
|
59
109
|
for (const h of whole) {
|
|
60
|
-
|
|
110
|
+
// The identity claim is PER HIT, not per tier: hits are ranked
|
|
111
|
+
// nearest-first, and grounding one below the bar under this tier's
|
|
112
|
+
// "near-identical" label would launder byte-overlap noise (observed:
|
|
113
|
+
// "merci" projecting through the unrelated near hit "meraih").
|
|
114
|
+
if (h.score < idBar)
|
|
115
|
+
break;
|
|
116
|
+
const own = read(ctx, h.id);
|
|
117
|
+
if (h.id === qId || restates(own)) {
|
|
61
118
|
const rev = ctx.store.prevFirst(h.id, hubBound(ctx));
|
|
62
119
|
const g = reverseContext(ctx, h.id, queryGist, rev);
|
|
63
120
|
if (g !== null) {
|
|
@@ -65,10 +122,7 @@ export async function recallByResonance(ctx, query, pre) {
|
|
|
65
122
|
? "perfect self-match — reverse recall to the sole predecessor"
|
|
66
123
|
: "perfect self-match — reverse recall to the best-resonating predecessor", nothing, STEP);
|
|
67
124
|
}
|
|
68
|
-
|
|
69
|
-
if (own.length > 0) {
|
|
70
|
-
return ground(own, "perfect self-match — the query IS this node", nothing, STEP);
|
|
71
|
-
}
|
|
125
|
+
continue;
|
|
72
126
|
}
|
|
73
127
|
const g = await project(ctx, h.id, queryGist);
|
|
74
128
|
if (g) {
|
|
@@ -76,8 +130,30 @@ export async function recallByResonance(ctx, query, pre) {
|
|
|
76
130
|
}
|
|
77
131
|
}
|
|
78
132
|
}
|
|
133
|
+
// The query-relative grounding fraction, shared by tiers 2–4 — gated on
|
|
134
|
+
// the FRACTION OF THE QUERY the grounding explains, not the raw cosine.
|
|
135
|
+
// Root gists are unit vectors, but their magnitudes are recoverable from
|
|
136
|
+
// the byte lengths (‖·‖ = √len under the linear fold):
|
|
137
|
+
// cos = shared/√(lenQ·lenG), so shared/lenQ = cos·√(lenG/lenQ).
|
|
138
|
+
// The raw cosine punished honest containment — a query fully inside a
|
|
139
|
+
// longer grounded answer scored √(lenQ/lenG) and was refused — and let a
|
|
140
|
+
// long answer sharing only scaffolding pass; the query-relative fraction
|
|
141
|
+
// measures exactly what the reach bar means: how much of THE QUERY the
|
|
142
|
+
// store accounts for.
|
|
143
|
+
// Chance similarity survives the length conversion AMPLIFIED: the same
|
|
144
|
+
// √(lenG/lenQ) factor that converts an honest shared fraction into a
|
|
145
|
+
// query-relative one multiplies the estimator/chance floor too, so a long
|
|
146
|
+
// stored form (√(lenG/lenQ) ≈ 10 at 100×) lifted a noise-level cosine past
|
|
147
|
+
// the reach bar and grounded pure gibberish (observed). Only the
|
|
148
|
+
// ABOVE-CHANCE part of the similarity is evidence of shared content —
|
|
149
|
+
// subtract the significance bar (3/√D, §8.3) before converting. Derived
|
|
150
|
+
// from the existing bars; never tuned.
|
|
151
|
+
const sig = significanceBar(ctx.store.D);
|
|
152
|
+
const reach = reachThreshold(ctx.space.maxGroup);
|
|
153
|
+
const fracOfQuery = (cos, otherLen) => Math.min(1, Math.max(0, cos - sig) *
|
|
154
|
+
Math.sqrt(otherLen / Math.max(1, query.length)));
|
|
79
155
|
// 2. Scaffolding-dominated.
|
|
80
|
-
if (top.score >=
|
|
156
|
+
if (top.score >= sig) {
|
|
81
157
|
const N = corpusN(ctx);
|
|
82
158
|
const minVote = consensusFloor(N);
|
|
83
159
|
// The committed points of attention ARE the shared climb's roots (same
|
|
@@ -86,44 +162,52 @@ export async function recallByResonance(ctx, query, pre) {
|
|
|
86
162
|
const forest = (await pre.attention()).roots;
|
|
87
163
|
if (forest.length > 0 && forest[0].vote >= minVote) {
|
|
88
164
|
const g = await project(ctx, forest[0].anchor, queryGist);
|
|
89
|
-
|
|
165
|
+
// The anchor cleared the consensus floor, but the floor prices the
|
|
166
|
+
// ANCHOR's evidence, not the projection's: a junk attractor can clear
|
|
167
|
+
// it and project a PIECE OF THE QUERY back at it (the observed
|
|
168
|
+
// "buenos días in English" → "English" fragment). A projection that
|
|
169
|
+
// is a proper byte-subspan of the query restates part of the question
|
|
170
|
+
// — never an answer (the same principle as `restates` above, extended
|
|
171
|
+
// to fragments). Genuine anchor groundings — longer than the query,
|
|
172
|
+
// or disjoint from it — pass untouched.
|
|
173
|
+
if (g && !(g.length < query.length && indexOf(query, g, 0) >= 0)) {
|
|
90
174
|
return ground(g, "scaffolding-dominated query — ground the consensus-climb anchor", [[forest[0].start, forest[0].end]], CONCEPT);
|
|
91
175
|
}
|
|
92
176
|
}
|
|
93
177
|
}
|
|
94
|
-
// 3. Last resort —
|
|
95
|
-
// explains, not the raw cosine. Root gists are unit vectors, but their
|
|
96
|
-
// magnitudes are recoverable from the byte lengths (‖·‖ = √len under the
|
|
97
|
-
// linear fold): cos = shared/√(lenQ·lenG), so shared/lenQ = cos·√(lenG/lenQ).
|
|
98
|
-
// The raw cosine punished honest containment — a query fully inside a
|
|
99
|
-
// longer grounded answer scored √(lenQ/lenG) and was refused — and let a
|
|
100
|
-
// long answer sharing only scaffolding pass; the query-relative fraction
|
|
101
|
-
// measures exactly what the reach bar means: how much of THE QUERY the
|
|
102
|
-
// store accounts for.
|
|
103
|
-
const fracOfQuery = (cos, otherLen) => Math.min(1, cos * Math.sqrt(otherLen / Math.max(1, query.length)));
|
|
178
|
+
// 3. Last resort — the nearest grounded whole-query hit, same gate.
|
|
104
179
|
for (const h of whole) {
|
|
105
180
|
const g = await project(ctx, h.id, queryGist);
|
|
106
181
|
if (g) {
|
|
107
182
|
if (fracOfQuery(cosine(queryGist, gistOf(ctx, g)), g.length) >=
|
|
108
|
-
|
|
183
|
+
reach) {
|
|
109
184
|
return ground(g, "last resort: the nearest grounded whole-query hit", [], STEP);
|
|
110
185
|
}
|
|
111
186
|
}
|
|
112
187
|
}
|
|
113
|
-
// The refusal/echo decision
|
|
114
|
-
//
|
|
115
|
-
//
|
|
116
|
-
//
|
|
117
|
-
// reach bar
|
|
118
|
-
//
|
|
119
|
-
//
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
188
|
+
// The refusal/echo decision. The echo returns a stored form's bytes AS
|
|
189
|
+
// the answer — a near-identity claim about the query — and identity-grade
|
|
190
|
+
// decisions are never made on an estimated score ("approximate scores may
|
|
191
|
+
// rank and propose; they may never decide", §6.2): the RaBitQ estimate
|
|
192
|
+
// overshooting the reach bar echoed a WRONG-entity neighbour ("capital of
|
|
193
|
+
// Zamunda?" echoed the Armenia fact, observed). The bytes are read
|
|
194
|
+
// anyway to be echoed, so the decision uses their EXACT fold: one river
|
|
195
|
+
// fold of the top hit, measured in the same query-relative,
|
|
196
|
+
// chance-corrected units as the tier above.
|
|
197
|
+
const topBytes = read(ctx, top.id);
|
|
198
|
+
const exact = topBytes.length > 0
|
|
199
|
+
? cosine(queryGist, gistOf(ctx, topBytes))
|
|
200
|
+
: 0;
|
|
201
|
+
if (fracOfQuery(exact, topBytes.length) < reach) {
|
|
123
202
|
return ground(null, "below reach threshold — nothing in the store relates to this query", [], 0);
|
|
124
203
|
}
|
|
204
|
+
// Echoing the query's own bytes back at it is not an echo of a RELATED
|
|
205
|
+
// form — it is the query restated, which answers nothing.
|
|
206
|
+
if (restates(topBytes)) {
|
|
207
|
+
return ground(null, "the nearest form IS the query itself — restating it answers nothing", [], 0);
|
|
208
|
+
}
|
|
125
209
|
// Honest echo.
|
|
126
|
-
return ground(
|
|
210
|
+
return ground(topBytes, "last resort: the nearest resonant form's own bytes (echo, not grounded)", [], 0, true);
|
|
127
211
|
}
|
|
128
212
|
// ── Pipeline mechanism ──────────────────────────────────────────────────────
|
|
129
213
|
export const recallMechanism = {
|
package/dist/src/mind/mind.d.ts
CHANGED
|
@@ -1,9 +1,10 @@
|
|
|
1
1
|
import { Vec } from "../vec.js";
|
|
2
2
|
import { Sema, Space } from "../sema.js";
|
|
3
3
|
import { Alphabet } from "../alphabet.js";
|
|
4
|
-
import { Grid } from "../geometry.js";
|
|
4
|
+
import { type FoldPyramid, Grid } from "../geometry.js";
|
|
5
5
|
import { BoundedMap, type Store } from "../store.js";
|
|
6
6
|
import { type MindConfig } from "../config.js";
|
|
7
|
+
import { type Canon } from "../canon.js";
|
|
7
8
|
import { GraphSearch, type Leaf, type Site } from "./graph-search.js";
|
|
8
9
|
import type { ComputedSpan, ExtensionHost } from "../extension.js";
|
|
9
10
|
export type { ComputedSpan, ExtensionHost };
|
|
@@ -19,6 +20,27 @@ export interface Response {
|
|
|
19
20
|
* no answer. */
|
|
20
21
|
provenance?: import("./pipeline.js").Provenance;
|
|
21
22
|
}
|
|
23
|
+
/** Serializable state of a conversation — can be saved and restored across
|
|
24
|
+
* sessions. The Mind never interprets the bytes; it only tracks their
|
|
25
|
+
* cumulative lengths so the caller can reconstruct turn boundaries later
|
|
26
|
+
* without inspecting content. */
|
|
27
|
+
export interface ConversationState {
|
|
28
|
+
/** The accumulated context bytes — raw concatenation of every turn's
|
|
29
|
+
* bytes in order. No separator is inserted; the boundary offsets
|
|
30
|
+
* ({@link boundaries}) tell the caller where each turn ends. */
|
|
31
|
+
context: Uint8Array;
|
|
32
|
+
/** Cumulative byte length after each completed turn. Sorted, strictly
|
|
33
|
+
* increasing, each {@code < context.length}. The first turn's length
|
|
34
|
+
* is `boundaries[0]`; the second turn starts at that offset, and so
|
|
35
|
+
* on. Empty for a single-turn or new conversation. */
|
|
36
|
+
boundaries: number[];
|
|
37
|
+
}
|
|
38
|
+
/** An active conversation handle. Opaque — interact through the Mind's
|
|
39
|
+
* conversation methods ({@link Mind.beginConversation},
|
|
40
|
+
* {@link Mind.respondTurn}, {@link Mind.endConversation}). */
|
|
41
|
+
export interface Conversation {
|
|
42
|
+
readonly id: number;
|
|
43
|
+
}
|
|
22
44
|
import type { AttentionRead, MindContext, Recognition } from "./types.js";
|
|
23
45
|
export interface MindOptions {
|
|
24
46
|
seed?: number;
|
|
@@ -34,6 +56,12 @@ export interface MindOptions {
|
|
|
34
56
|
mechanisms?: import("./pipeline-mechanism.js").PipelineMechanism[];
|
|
35
57
|
/** Factories that receive the {@link ExtensionHost} and return mechanisms. */
|
|
36
58
|
mechanismFactories?: ((host: import("../extension.js").ExtensionHost) => import("./pipeline-mechanism.js").PipelineMechanism)[];
|
|
59
|
+
/** Content canonicalizer applied to EVERY response (any modality) for
|
|
60
|
+
* equivalence-class resolution — see src/canon.ts. Text entry points
|
|
61
|
+
* ({@link Mind.respondText}, {@link Mind.respondTurnText}) inject the
|
|
62
|
+
* Unicode text canonicalizer automatically when this is unset; pass
|
|
63
|
+
* `false` to disable canonical resolution everywhere. */
|
|
64
|
+
canon?: Canon | false;
|
|
37
65
|
}
|
|
38
66
|
export declare class Mind implements MindContext {
|
|
39
67
|
readonly space: Space;
|
|
@@ -46,16 +74,27 @@ export declare class Mind implements MindContext {
|
|
|
46
74
|
readonly mechanisms: import("./pipeline-mechanism.js").PipelineMechanism[];
|
|
47
75
|
/** The live rationale tracer for the inference currently in flight, or null. */
|
|
48
76
|
trace: Rationale | null;
|
|
49
|
-
/**
|
|
50
|
-
* {@link
|
|
51
|
-
*
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
/**
|
|
56
|
-
|
|
57
|
-
|
|
77
|
+
/** The content canonicalizer for the response in flight — see
|
|
78
|
+
* {@link MindContext.canon}. Injected per response by the modality entry
|
|
79
|
+
* point; null when the response carries no equivalence. */
|
|
80
|
+
canon: Canon | null;
|
|
81
|
+
/** Per-response canonical-resolution memo — see {@link MindContext.canonMemo}. */
|
|
82
|
+
canonMemo: Map<string, number | null> | null;
|
|
83
|
+
/** The Mind-level canon option: a canonicalizer to use for EVERY response,
|
|
84
|
+
* `false` to disable canonical resolution, or null to let each entry
|
|
85
|
+
* point decide (text entry points inject {@link textCanon}). */
|
|
86
|
+
private _canonOpt;
|
|
87
|
+
/** Memo of the consensus climb — content-keyed. See {@link MindContext.climbMemo}. */
|
|
88
|
+
climbMemo: Map<string, Map<string, AttentionRead>> | null;
|
|
89
|
+
/** Memo of recognise() — content-keyed. See {@link MindContext.recogniseMemo}. */
|
|
90
|
+
recogniseMemo: Map<string, Recognition> | null;
|
|
91
|
+
/** Memo of perceive() — content-keyed. See {@link MindContext.perceiveMemo}. */
|
|
58
92
|
perceiveMemo: Map<string, import("../sema.js").Sema> | null;
|
|
93
|
+
/** Subtree-resolution cache. See {@link MindContext._resolvedSubtrees}. */
|
|
94
|
+
_resolvedSubtrees: WeakMap<import("../sema.js").Sema, {
|
|
95
|
+
id: number;
|
|
96
|
+
len: number;
|
|
97
|
+
}> | null;
|
|
59
98
|
/** The perceived gist of the query currently being answered. Set by `think`
|
|
60
99
|
* before the graph search runs; `chooseNext` consults it as a gate (a null
|
|
61
100
|
* guide means no query is in flight, so structural walkers keep plain
|
|
@@ -72,9 +111,11 @@ export declare class Mind implements MindContext {
|
|
|
72
111
|
* {@link MindContext._gistCache}. 32 MB ≈ 8K gists at D=1024; hub
|
|
73
112
|
* candidate sets (√N at most) fit comfortably and recur across queries. */
|
|
74
113
|
_gistCache: BoundedMap<number, Vec>;
|
|
75
|
-
_depositTrees: BoundedMap<string,
|
|
114
|
+
_depositTrees: BoundedMap<string, FoldPyramid>;
|
|
76
115
|
_depositLens: Set<number>;
|
|
77
116
|
_internIds: WeakMap<Sema, number>;
|
|
117
|
+
private _nextConvId;
|
|
118
|
+
private _conversations;
|
|
78
119
|
/** Canonical node id of a byte span. Required by GraphSearchHost & MindContext. */
|
|
79
120
|
resolve(bytes: Uint8Array): number | null;
|
|
80
121
|
recogniseSpan(bytes: Uint8Array): {
|
|
@@ -106,15 +147,85 @@ export declare class Mind implements MindContext {
|
|
|
106
147
|
* memos. Paired with {@link endResponse}; the ONE place this state is
|
|
107
148
|
* created, so adding a memo cannot forget its reset. */
|
|
108
149
|
private beginResponse;
|
|
150
|
+
/** The canonicalizer a response should carry: the Mind-level option when
|
|
151
|
+
* set (or none when explicitly disabled), else the entry point's own
|
|
152
|
+
* default — text entry points pass {@link textCanon}, binary ones null. */
|
|
153
|
+
private _canonFor;
|
|
109
154
|
/** Close one response's transient state — every per-response field, incl.
|
|
110
155
|
* the edge guide/choices `think` sets mid-flight. */
|
|
111
156
|
private endResponse;
|
|
157
|
+
/** Shared response core — the one path from bytes to voiced answer.
|
|
158
|
+
* `respond` calls this directly; `respondTurn` has its own path
|
|
159
|
+
* with conversation-persistent memos and incremental perception. */
|
|
160
|
+
private _respondImpl;
|
|
112
161
|
respond(input: Input, inspectRationale?: InspectRationale): Promise<Response>;
|
|
113
162
|
/** Text view of {@link respond}. NUL bytes (0x00) are stripped before
|
|
114
163
|
* decoding — they are structural padding in text answers. LOSSY for a
|
|
115
164
|
* binary answer that legitimately contains NULs: use {@link respond} and
|
|
116
|
-
* read `bytes` directly for binary/grid modalities.
|
|
165
|
+
* read `bytes` directly for binary/grid modalities.
|
|
166
|
+
*
|
|
167
|
+
* Injects the TEXT canonicalizer (src/canon.ts) so resolution treats
|
|
168
|
+
* every character variation of the same text — case, width, whitespace —
|
|
169
|
+
* as one form, provided the store's canon index is built
|
|
170
|
+
* ({@link buildCanonIndex}). */
|
|
117
171
|
respondText(input: string, inspectRationale?: InspectRationale): Promise<string>;
|
|
172
|
+
/** Begin a new conversation, optionally restoring from a previously-saved
|
|
173
|
+
* {@link ConversationState}. The returned handle is required for
|
|
174
|
+
* {@link respondTurn} and {@link endConversation}.
|
|
175
|
+
*
|
|
176
|
+
* Conversations are independent — a Mind can manage several concurrently.
|
|
177
|
+
* Each tracks the fold pyramid (accumulated internal processing) and
|
|
178
|
+
* turn-boundary offsets; the geometry never inspects content to guess
|
|
179
|
+
* where one turn ends and the next begins. */
|
|
180
|
+
beginConversation(state?: ConversationState): Conversation;
|
|
181
|
+
/** End a conversation, releasing its internal resources (accumulated
|
|
182
|
+
* context, boundary offsets, and the fold-pyramid cache). Idempotent. */
|
|
183
|
+
endConversation(conv: Conversation): void;
|
|
184
|
+
/** The current serialisable state of an active conversation. Save this
|
|
185
|
+
* to resume the conversation later via {@link beginConversation}. */
|
|
186
|
+
conversationState(conv: Conversation): ConversationState | null;
|
|
187
|
+
/** Append a turn to a conversation's accumulated context WITHOUT
|
|
188
|
+
* responding — raw byte append plus a boundary offset, never a
|
|
189
|
+
* separator; the fold pyramid advances by O(turn).
|
|
190
|
+
*
|
|
191
|
+
* This is the primitive for turns the Mind should hear but not answer:
|
|
192
|
+
* replaying a transcript, feeding the OTHER speaker's line in a
|
|
193
|
+
* prediction harness, or restoring context piecewise. {@link
|
|
194
|
+
* respondTurn} = addTurn + think + its own reply appended the same way. */
|
|
195
|
+
addTurn(conv: Conversation, turn: Input): ConversationState;
|
|
196
|
+
/** Grow a conversation's accumulated context by one turn's bytes — raw
|
|
197
|
+
* append plus a boundary offset, pyramid advanced by O(turn), the grown
|
|
198
|
+
* context's tree seeded into the conversation's perceive memo. The ONE
|
|
199
|
+
* place a context grows ({@link addTurn} and {@link respondTurn} both
|
|
200
|
+
* come through here), so the append semantics cannot drift. */
|
|
201
|
+
private _growContext;
|
|
202
|
+
/** Process one turn of a conversation.
|
|
203
|
+
*
|
|
204
|
+
* `turn` is the raw input for the latest turn — its bytes are appended
|
|
205
|
+
* to the accumulated context directly (raw concatenation). The Mind
|
|
206
|
+
* tracks the byte offset where each turn ends; no separator is ever
|
|
207
|
+
* inserted or inspected.
|
|
208
|
+
*
|
|
209
|
+
* Returns the response AND the updated {@link ConversationState} so the
|
|
210
|
+
* caller can persist it. The conversation handle's internal state is
|
|
211
|
+
* updated in place — the returned state is a snapshot for storage.
|
|
212
|
+
*
|
|
213
|
+
* SINGLE FLIGHT: at most one respondTurn may be in flight per Mind. The
|
|
214
|
+
* conversation's memo caches are swapped into the Mind-level per-response
|
|
215
|
+
* pointers for the duration of the turn, so a concurrently-running
|
|
216
|
+
* respond()/respondTurn() on the SAME Mind would interleave state.
|
|
217
|
+
* Different Minds (or sequential awaits, as in every test) are safe. */
|
|
218
|
+
respondTurn(conv: Conversation, turn: Input, inspectRationale?: InspectRationale): Promise<{
|
|
219
|
+
response: Response;
|
|
220
|
+
state: ConversationState;
|
|
221
|
+
}>;
|
|
222
|
+
/** Text view of {@link respondTurn}. See {@link respondText} for the
|
|
223
|
+
* NUL-stripping caveat. For binary or grid turns use {@link respondTurn}
|
|
224
|
+
* directly — this is a text-only convenience, like {@link respondText}. */
|
|
225
|
+
respondTurnText(conv: Conversation, turn: string, inspectRationale?: InspectRationale): Promise<{
|
|
226
|
+
response: string;
|
|
227
|
+
state: ConversationState;
|
|
228
|
+
}>;
|
|
118
229
|
embedding(input: Input): Promise<Vec | null>;
|
|
119
230
|
/** Kinship note: the vector arm below is a miniature of recall's tier 3
|
|
120
231
|
* (resonate → reach gate → read out the nearest form's bytes) — the
|
|
@@ -141,6 +252,21 @@ export declare class Mind implements MindContext {
|
|
|
141
252
|
* (default 2 — structural bridges)
|
|
142
253
|
* @returns number of nodes added to the content index */
|
|
143
254
|
repairContentIndex(minParents?: number): Promise<number>;
|
|
255
|
+
/** Build (or incrementally refresh) the store's canonical-form index: for
|
|
256
|
+
* every content-bearing node, record the hash of its CANONICAL key so
|
|
257
|
+
* resolution can find stored forms across surface variation (case, width,
|
|
258
|
+
* whitespace — whatever `canon` equates; see src/canon.ts).
|
|
259
|
+
*
|
|
260
|
+
* Incremental and idempotent: the last indexed node id is remembered in
|
|
261
|
+
* store meta (`canon.upto`), so a refresh after further training scans
|
|
262
|
+
* only the new rows. Run once after training, and again after ingests —
|
|
263
|
+
* the same operational shape as {@link repairContentIndex}.
|
|
264
|
+
*
|
|
265
|
+
* @param canon the canonicalizer to index under — MUST be the same one
|
|
266
|
+
* queries will carry (text queries carry {@link textCanon}
|
|
267
|
+
* unless the Mind was constructed with its own)
|
|
268
|
+
* @returns number of index rows added */
|
|
269
|
+
buildCanonIndex(canon?: Canon): Promise<number>;
|
|
144
270
|
save(): Promise<Uint8Array>;
|
|
145
271
|
static load(snapshot: Uint8Array, store: Store): Promise<Mind>;
|
|
146
272
|
static loadFromStore(store: Store): Promise<Mind>;
|