@hviana/sema 0.1.0 → 0.1.2

package/dist/src/ingest-cache.d.ts

@@ -0,0 +1,41 @@
+import { Mind } from "./mind/index.js";
+import { type Sema } from "./sema.js";
+import type { Input } from "./mind/index.js";
+/**
+ * An ingest cache layered over a Mind.
+ *
+ * Usage (drop-in for mind.ingest in training loops):
+ *   const ci = new CachedIngest(mind);
+ *   await ci.ingest(ctx, cont);       // caches both sides
+ *   await ci.ingest("some text");     // caches the single text
+ *   await ci.ingest([[ctx, cont]]);   // array form
+ *
+ * All other Mind methods (respond, save, …) are accessed on `ci.mind`.
+ */
+export declare class CachedIngest {
+    readonly mind: Mind;
+    /** One unified, content-addressed memo of interned inputs.  Each value holds
+     *  Float32 root vector (D·4 bytes) plus a small id array, so the byte budget
+     *  is spent on the inputs most worth remembering (LRU). */
+    private _memo;
+    hits: number;
+    misses: number;
+    constructor(mind: Mind);
+    private newMemo;
+    private keyOf;
+    ingest(input: Input | (Input | [Input, Input])[], second?: Input): Promise<(Sema & {
+        id: number;
+    }) | undefined>;
+    /** Resolve an input to its interned nodes, from the memo when its content has
+     *  been seen before, else by the Mind's OWN untracked {@link deposit}
+     *  (perceive + intern + sub-span/containment indexing — identical to the
+     *  direct path; those writes are durable and idempotent, so a later memo hit
+     *  legitimately skips them).  This is the single expensive operation both
+     *  ingest paths share; caching it here is what makes every repeated input —
+     *  in any role — cheap. */
+    private resolveInput;
+    private ingestPair;
+    private ingestOne;
+    clear(): void;
+    get size(): number;
+}

package/dist/src/ingest-cache.js

@@ -0,0 +1,161 @@
+// ingest-cache.ts — ingest-level cache for Mind.
+//
+// Wraps a Mind and memoises the expensive, content-determined half of ingestion:
+// turning ONE input (a context, a continuation, or a bare experience) into its
+// interned graph nodes.  Perception (riverFold) and interning are a PURE function
+// of the input bytes — the same bytes always cut the same way and hash-cons to
+// the same node ids — so the result can be cached by content and reused wherever
+// those same bytes appear again, in ANY role.
+//
+// This is why the cache keys on the single input, not on the (context,
+// continuation) pair: real training corpora repeat content at the INPUT level far
+// more than at the pair level — a shared system/tool preamble in front of every
+// row, a common answer ("Yes.", a refusal) under many questions, an accumulated
+// dialogue prefix, an exactly-duplicated row.  A pair-level key sees none of these
+// (the pair as a whole is almost always unique); the input-level key catches them
+// all, and a context that recurs as a continuation (or vice versa) hits too.
+//
+// On a cache hit only the cheap relational writes are replayed — the edge and the
+// two halo pours for a pair, or the part-chain links for a single — while the
+// perceive + internTree work (the dominant cost, and on a large store the dominant
+// VECTOR-INDEX cost too) is skipped entirely.  The interned node ids are stable
+// across eviction: a re-perceive of evicted content reproduces the same ids by
+// the store's hash-cons, so correctness never depends on a hit.
+//
+// DEPOSITION IS THE MIND'S OWN: every miss goes through the same
+// {@link deposit} the direct Mind.ingest path uses — perception, interning,
+// sub-span/window indexing, durable CONTAINMENT edges, whole-stream flat
+// branch, changed-node tracking against the Mind's `_prevSeen` — so a store
+// trained through this cache is structurally IDENTICAL to one trained by
+// calling mind.ingest() directly (this class used to re-implement the deposit
+// and drifted: no indexSubSpans/addContainer, wrong part-link stride).  Only
+// the memo layer lives here.  Behaviour is identical to mind.ingest() (links
+// and halos are reinforced on every call), just faster when content repeats.
+import { deposit, dispatchIngest, ingestOne as depositOne, } from "./mind/learning.js";
+import { bindSeat, companySignature } from "./sema.js";
+import { BoundedMap } from "./store.js";
+/**
+ * An ingest cache layered over a Mind.
+ *
+ * Usage (drop-in for mind.ingest in training loops):
+ *   const ci = new CachedIngest(mind);
+ *   await ci.ingest(ctx, cont);       // caches both sides
+ *   await ci.ingest("some text");     // caches the single text
+ *   await ci.ingest([[ctx, cont]]);   // array form
+ *
+ * All other Mind methods (respond, save, …) are accessed on `ci.mind`.
+ */
+export class CachedIngest {
+    mind;
+    /** One unified, content-addressed memo of interned inputs.  Each value holds
+     *  Float32 root vector (D·4 bytes) plus a small id array, so the byte budget
+     *  is spent on the inputs most worth remembering (LRU). */
+    _memo;
+    hits = 0;
+    misses = 0;
+    constructor(mind) {
+        this.mind = mind;
+        this._memo = this.newMemo();
+    }
+    newMemo() {
+        return new BoundedMap(this.mind.cfg.store.ingestCacheBytes, (c) => c.rootV.byteLength + c.partIds.length * 4 + 8 + c.keyBytes);
+    }
+    // ── cache key ───────────────────────────────────────────────────────────
+    // Content key for an input, or null when the input cannot be keyed reliably
+    // — in which case it BYPASSES the cache (recomputed every time) rather than
+    // risk a collision returning the wrong nodes.  A string is its own key; raw
+    // bytes hash by FNV-1a (+ length, so a hash collision still needs an equal
+    // length to alias); a Grid / Grid[] has no cheap stable key, so it bypasses.
+    keyOf(input) {
+        if (typeof input === "string")
+            return "S:" + input;
+        if (input instanceof Uint8Array) {
+            // FNV-1a — no spread onto the call stack, safe on large binary inputs.
+            let h = 0x811c9dc5 >>> 0;
+            for (let i = 0; i < input.length; i++) {
+                h ^= input[i];
+                h = Math.imul(h, 0x01000193) >>> 0;
+            }
+            return "B:" + (h >>> 0).toString(16) + ":" + input.length;
+        }
+        return null; // Grid / Grid[] — bypass the cache
+    }
+    // ── public API ────────────────────────────────────────────────────────
+    async ingest(input, second) {
+        // One shape-reading for both ingest paths — see {@link dispatchIngest}.
+        return dispatchIngest(input, second, (i) => this.ingestOne(i), (a, b) => this.ingestPair(a, b));
+    }
+    // ── the one cached step: perceive + intern + index ONE input ──────────────
+    /** Resolve an input to its interned nodes, from the memo when its content has
+     *  been seen before, else by the Mind's OWN untracked {@link deposit}
+     *  (perceive + intern + sub-span/containment indexing — identical to the
+     *  direct path; those writes are durable and idempotent, so a later memo hit
+     *  legitimately skips them).  This is the single expensive operation both
+     *  ingest paths share; caching it here is what makes every repeated input —
+     *  in any role — cheap. */
+    async resolveInput(input) {
+        const key = this.keyOf(input);
+        if (key !== null) {
+            const cached = this._memo.get(key);
+            if (cached) {
+                this.hits++;
+                return cached;
+            }
+        }
+        this.misses++;
+        // The full deposit, untracked (a continuation must not break the tracked
+        // deposit chain).  `partIds` are the root's immediate children, already
+        // interned by the deposit — free to capture.  Only the root vector and
+        // ids are retained; the tree is GC'd.
+        const { tree, rootId, ids } = await deposit(this.mind, input, false);
+        const partIds = tree.kids
+            ? tree.kids.map((k) => ids.get(k))
+            : [rootId];
+        const entry = {
+            rootV: tree.v,
+            rootId,
+            partIds,
+            // ~2 bytes per UTF-16 code unit — the key is retained by the map.
+            keyBytes: key !== null ? key.length * 2 : 0,
+        };
+        if (key !== null)
+            this._memo.set(key, entry);
+        return entry;
+    }
+    // ── pair ingest ───────────────────────────────────────────────────────
+    async ingestPair(ctx, cont) {
+        // The CONTEXT side is TRACKED (it continues the Mind's deposit chain and
+        // is the growing, unique-per-turn side — a cache miss anyway), via the
+        // same {@link deposit} the direct path uses; the continuation resolves
+        // through the memo.
+        const c = await deposit(this.mind, ctx, true);
+        const b = await this.resolveInput(cont);
+        // EDGE on the FULL context (disambiguates a shared pivot). HALO only on the
+        // CHANGED NODES — the coherent new subtrees this turn adds — so a recurring
+        // shared prefix never collects halo mass. (See Mind.ingestPair.)
+        await this.mind.store.link(c.rootId, b.rootId);
+        // Halos pour company SIGNATURES (identity), not gists (content) — see
+        // companySignature in sema.ts.
+        const contSeat = bindSeat(this.mind.space, companySignature(this.mind.space, b.rootId), 1);
+        for (const part of c.changed) {
+            const partId = c.ids.get(part);
+            await this.mind.store.pourHalo(partId, contSeat);
+            await this.mind.store.pourHalo(b.rootId, bindSeat(this.mind.space, companySignature(this.mind.space, partId), 0));
+        }
+    }
+    // ── one-side ingest ───────────────────────────────────────────────────
+    async ingestOne(input) {
+        // A bare experience is the direct path verbatim — tracked deposit, root
+        // marked a resonance target, part chain linked at the maxGroup stride.
+        // (The tracked side is a memo miss by design, so there is nothing for the
+        // cache to add here.)
+        return depositOne(this.mind, input);
+    }
+    // ── helpers ───────────────────────────────────────────────────────────
+    clear() {
+        this._memo = this.newMemo();
+    }
+    get size() {
+        return this._memo.size;
+    }
+}

package/dist/src/mind/articulation.d.ts

@@ -0,0 +1,6 @@
+import type { MindContext } from "./types.js";
+/** Re-voice an answer in the asker's own words.  For each recognised form
+ *  in the answer, find a concept-sibling in the query (by halo resonance)
+ *  and substitute the asker's wording.  The search's own cover mechanism
+ *  splices the substitutes into the answer exactly where the forms sit. */
+export declare function articulate(ctx: MindContext, answer: Uint8Array, query: Uint8Array | null): Promise<Uint8Array>;

package/dist/src/mind/articulation.js

@@ -0,0 +1,99 @@
+// articulation.ts — re-voice an answer in the asker's wording (Section 5).
+//
+//   articulate — substitute answer forms with the asker's synonyms,
+//                using concept (halo) resonance to match the voices.
+import { spliceAll } from "./types.js";
+import { recognise } from "./recognition.js";
+import { contains } from "./traverse.js";
+import { bestHaloMate } from "./match.js";
+import { coverSequence } from "../derive/src/index.js";
+import { rItem, rNode, traceDerivation } from "./trace.js";
+/** Re-voice an answer in the asker's own words.  For each recognised form
+ *  in the answer, find a concept-sibling in the query (by halo resonance)
+ *  and substitute the asker's wording.  The search's own cover mechanism
+ *  splices the substitutes into the answer exactly where the forms sit. */
+export async function articulate(ctx, answer, query) {
+    if (!query || query.length < 2 || answer.length < 2)
+        return answer;
+    const store = ctx.store;
+    const t = ctx.trace?.enter("articulate", [
+        rItem(answer, "answer"),
+        rItem(query, "query"),
+    ]);
+    const keep = (note) => {
+        t?.done([rItem(answer, "answer")], note);
+        return answer;
+    };
+    const qCandidates = [];
+    for (const s of recognise(ctx, query).sites) {
+        // payload < 0 is a SINGLE-BYTE leaf (byte leaves occupy the implicit
+        // negative id range −256…−1 — see store.ts).  A one-byte form's halo is
+        // promiscuous (it keeps company with everything it ever neighboured), so
+        // admitting it as a voice or a revoicing target lets e.g. the digit "4"
+        // be "re-voiced" into an unrelated digit the query happens to contain.
+        // Articulation therefore operates on multi-byte forms only.
+        if (s.payload < 0)
+            continue;
+        const h = store.halo(s.payload);
+        if (!h)
+            continue;
+        qCandidates.push({
+            start: s.start,
+            end: s.end,
+            payload: {
+                halo: h,
+                bytes: query.slice(s.start, s.end),
+                node: s.payload,
+            },
+        });
+    }
+    if (qCandidates.length === 0) {
+        return keep("no asker concept to revoice — answer unchanged");
+    }
+    if (coverSequence(query.length, qCandidates).spans.length === 0) {
+        return keep("no asker concept to revoice — answer unchanged");
+    }
+    const voices = qCandidates.map((s) => s.payload);
+    const ans = recognise(ctx, answer);
+    const substitutions = new Map();
+    for (const s of ans.sites) {
+        // Same single-byte-leaf exclusion as the query loop above.
+        if (s.payload < 0)
+            continue;
+        const h = store.halo(s.payload);
+        if (!h)
+            continue;
+        const found = bestHaloMate(ctx, h, voices, (v) => v.halo);
+        if (!found)
+            continue;
+        const voice = found.item;
+        if (voice.node === s.payload || contains(ctx, voice.node, s.payload)) {
+            continue;
+        }
+        substitutions.set(s.payload, voice.bytes);
+    }
+    if (substitutions.size === 0) {
+        return keep("no answer form is a synonym of an asker concept");
+    }
+    ctx.trace?.step("substitute", [...substitutions.keys()].map((n) => rNode(ctx, n, "answer-form")), [...substitutions.values()].map((b) => rItem(b, "asker-voice")), `revoice ${substitutions.size} answer form(s) in the asker's own words (synonym splice)`);
+    const voicedSites = ans.sites.filter((s) => substitutions.has(s.payload));
+    const tArtCover = ctx.trace?.enter("cover", [
+        ...voicedSites.map((s) => rItem(answer.subarray(s.start, s.end), "form", s.payload, [
+            s.start,
+            s.end,
+        ])),
+    ]);
+    const solved = ctx.search.cover(answer.length, voicedSites, new Map(), ans.leaves, ans.splits, substitutions, undefined, undefined, ctx.trace ? (steps) => traceDerivation(ctx, steps) : undefined);
+    const segs = solved && solved.segs;
+    tArtCover?.done(segs === null
+        ? []
+        : segs.map((s) => rItem(s.bytes, s.rec ? "voiced" : "kept", s.node, [s.i, s.j])), segs === null
+        ? "the revoiced cover did not compose"
+        : "lightest derivation of the revoiced answer");
+    const result = segs && spliceAll(segs);
+    if (result) {
+        t?.done([rItem(result, "voiced", ctx.store.findLeaf(result.subarray(0, 1)) ?? undefined)], "splice the asker's wording into the answer where the same concept appears");
+        return result;
+    }
+    return keep("the revoiced cover did not compose — answer unchanged");
+}

package/dist/src/mind/attention.d.ts

@@ -0,0 +1,72 @@
+import type { DerivationStep } from "./graph-search.js";
+import type { AncestorReach, Attention, AttentionRead, DFMode, MindContext, Region, RegionVote, SaturationInfo } from "./types.js";
+/** Climb the query's perceived byte regions up the structural DAG via
+ *  resonance, pool the evidence, and return only the ROOT points of
+ *  attention — those that cleared commitVotes' significance floor. */
+export declare function climbAttention(ctx: MindContext, query: Uint8Array, k: number, mode?: DFMode): Promise<Attention[]>;
+/** Full read-out of one consensus climb: both the roots (dominant points of
+ *  attention) and the entire ranked list.  Cached via ctx.climbMemo when
+ *  ctx.trace is null. */
+export declare function climbAttentionAll(ctx: MindContext, query: Uint8Array, k: number, mode?: DFMode): Promise<AttentionRead>;
+export declare function computeAttention(ctx: MindContext, query: Uint8Array, k: number, mode: DFMode): Promise<AttentionRead>;
+export declare function collectRegions(ctx: MindContext, query: Uint8Array): Region[];
+export declare function voteRegions(ctx: MindContext, query: Uint8Array, regions: readonly Region[], k: number, mode: DFMode, N: number, reachMemo?: Map<number, AncestorReach>): Promise<{
+    votes: RegionVote[];
+    saturated: boolean[];
+    voters: Array<{
+        id: number;
+        score: number;
+        w: number;
+    } | null>;
+}>;
+/** The consensus vote as EVIDENCE POOLING, not shortest path: each surviving
+ *  region is an axiom; it contributes to every root it climbed to (or, for a
+ *  terminal answer node, to the contexts that lead to it) by a `combine:
+ *  "sum"` rule, so independent regions corroborating the same anchor ADD
+ *  rather than compete to be the cheapest route (see {@link Rule.combine} in
+ *  derive/src/deduction.ts).  Run through the very engine {@link
+ *  GraphSearch} covers with — `lightestDerivation` — so a pooled-evidence
+ *  decision is, like a followed edge or a spliced connector, one weighted
+ *  rule of the SAME deduction system, not a separate hand-rolled tally that
+ *  merely logs alongside it.  `votesIdf`/`support` are the same two
+ *  read-outs {@link commitVotes} always gated on; only how they accumulate
+ *  changed. */
+export declare function poolVotes(ctx: MindContext, regionVotes: readonly RegionVote[], sat: SaturationInfo, N: number): {
+    votes: Map<number, number>;
+    votesIdf: Map<number, number>;
+    support: Map<number, {
+        start: number;
+        end: number;
+        w: number;
+    }>;
+    steps: DerivationStep[];
+};
+export declare function commitVotes(ctx: MindContext, pooled: {
+    votes: Map<number, number>;
+    votesIdf: Map<number, number>;
+    support: Map<number, {
+        start: number;
+        end: number;
+        w: number;
+    }>;
+    steps: DerivationStep[];
+}, sat: SaturationInfo, regions: readonly Region[], regionVoter: ReadonlyArray<{
+    id: number;
+    score: number;
+    w: number;
+} | null>, N: number): AttentionRead;
+export declare function detectSaturated(ctx: MindContext, regions: ReadonlyArray<{
+    start: number;
+    end: number;
+    chunk?: boolean;
+}>, saturated: ReadonlyArray<boolean>): SaturationInfo;
+export declare function canonicalChunkId(ctx: MindContext, regionBytes: Uint8Array, N: number, reachMemo?: Map<number, AncestorReach>): number | null;
+export declare function naturalBreak(votes: number[]): number;
+export declare function traceAttention(ctx: MindContext, regions: ReadonlyArray<{
+    start: number;
+    end: number;
+}>, regionVoter: ReadonlyArray<{
+    id: number;
+    score: number;
+    w: number;
+} | null>, roots: ReadonlyArray<Attention>, steps?: ReadonlyArray<DerivationStep>): void;