agentfootprint 8.7.0 → 8.9.0

package/dist/adapters/memory/sqliteVector.js

@@ -0,0 +1,875 @@
+"use strict";
+/**
+ * adapters/memory/sqliteVector — a corpus in a file, so you embed it once.
+ *
+ * `InMemoryStore` is a `Map`, and its cost is the one nobody notices until the
+ * bill arrives: **restart the process and the whole corpus is re-embedded.**
+ * That is fine for three documents and absurd for ten thousand. The step up was
+ * "bring a vector database", and the step between those two — *one machine, one
+ * file, nothing to install* — was a store every consumer had to write for
+ * themselves. This is that store.
+ *
+ * ── The two-phase cost model this exists to fix ─────────────────────────────
+ * **Index time** embeds the corpus. It happens once, in its own process, and
+ * its cost scales with the size of your documents. **Query time** embeds one
+ * thing — the user's question — per retrieval, and its cost scales with
+ * traffic. A 10,000-chunk corpus is 10,000 embeddings *once* and one embedding
+ * per question thereafter. With a `Map` it is 10,000 embeddings *per restart*.
+ * That is the whole argument for a file.
+ *
+ * ── Exact search, and the ceiling said out loud ─────────────────────────────
+ * Vectors live in SQLite as `Float32Array` blobs. On the first search of a
+ * namespace they are hydrated into ONE resident `Float32Array` matrix,
+ * normalised, and every later query is an exact dot product across it. There is
+ * **no approximate index and no pretence of one**: this returns the true top-K,
+ * or it does not answer.
+ *
+ * Measured against THIS implementation on Node 22.16, Apple silicon, one
+ * namespace, median of five queries:
+ *
+ * | corpus | query | resident matrix | file | first search (hydration) |
+ * |---|---|---|---|---|
+ * | 10,000 × 384-d | 6 ms | 15 MB | 21 MB | 45 ms |
+ * | 50,000 × 384-d | 31 ms | 77 MB | 105 MB | 251 ms |
+ * | 100,000 × 384-d | 65 ms | 154 MB | 211 MB | 939 ms |
+ * | 10,000 × 1536-d | 16 ms | 61 MB | 83 MB | 122 ms |
+ * | 50,000 × 1536-d | 89 ms | 307 MB | 413 MB | **5.7 s** |
+ *
+ * **The documented ceiling is 50,000 chunks.** Below it, every query is under
+ * 100 ms at every embedder this library ships and the resident matrix is under
+ * ~300 MB. It degrades linearly and predictably to about 100,000. Above that,
+ * or when the process cannot hold the matrix, move to a managed vector database
+ * — `MemoryStore` is the seam, and nothing else in your code changes. For scale
+ * intuition: 50,000 chunks at ~1,000 characters is roughly 50 MB of text, on
+ * the order of 25,000 pages.
+ *
+ * **Hydration is the number to plan around, not the query.** Steady-state
+ * search is fast everywhere in that table; reading the vectors off disk the
+ * FIRST time is what costs, and at 50,000 × 1536 it is 5.7 seconds. Paid
+ * lazily, that lands on whoever asks the first question after a deploy. Call
+ * {@link SqliteVectorStore.warm} at boot to pay it somewhere you chose — see
+ * that method. Smaller vectors are dramatically cheaper here: 384 dimensions
+ * hydrates 100,000 chunks in under a second, which is one more reason a
+ * 384-dimension embedder is the better default for a corpus this size.
+ *
+ * A loadable extension (sqlite-vec) was measured against this and deliberately
+ * NOT taken: it is roughly 2× faster at these sizes, and costs a native binary
+ * on a five-platform matrix (no musl, no Windows/arm64) plus a pre-1.0
+ * dependency. Two times, at sizes where we are already under 100 ms, does not
+ * buy that.
+ *
+ * ── One process on one machine ──────────────────────────────────────────────
+ * The same ceiling `sqliteSessions` states, for the same reason. It survives a
+ * restart, a crash, a deploy. It is NOT distributed. WAL gives one writer and
+ * many readers at once; a second writer waits up to `busyTimeoutMs` and then
+ * fails loudly rather than queueing forever.
+ *
+ * ── Zero dependencies, and the version floor that buys ──────────────────────
+ * SQLite is *inside Node* — no install, no native build, no peer dependency.
+ * The price is a version floor this package does not otherwise have, so the
+ * module is loaded when you actually construct a store and its absence is
+ * refused by name. There is deliberately **no fallback to memory**: a corpus
+ * that silently forgot every document on restart looks, from the outside,
+ * exactly like a corpus that was never built.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.sqliteVectorStore = exports.EmbedderMismatchError = exports.UnreadableIndexFileError = void 0;
+const index_js_1 = require("../../memory/identity/index.js");
+const lazyRequire_js_1 = require("../../lib/lazyRequire.js");
+const sqliteUnavailable_js_1 = require("../../lib/sqliteUnavailable.js");
+// ─── The refusals ────────────────────────────────────────────────────
+/**
+ * Raised when the file exists but this runtime cannot use it as a vector index.
+ *
+ * The law `sqliteSessions` states for a session file, one domain over: **an
+ * unreadable index and an empty one are different facts, and only one of them
+ * is safe to answer with "no matches".** A store that opened a corrupt file as
+ * an empty database would answer every question from the model's own weights
+ * and log nothing.
+ *
+ * `problem` is the fact to branch on:
+ *
+ *  - `'cannot-open'` — not a SQLite database, or not readable.
+ *  - `'not-our-schema'` — a database whose `af_vectors` table is somebody
+ *    else's table of that name. Point the store at its own file.
+ *  - `'newer-schema'` — written by a newer agentfootprint than this one.
+ */
+class UnreadableIndexFileError extends Error {
+    code = 'ERR_UNREADABLE_INDEX_FILE';
+    /** The file that was refused. */
+    file;
+    /** Which of the three cases this is. */
+    problem;
+    constructor(file, problem, detail) {
+        super(`[memory] the vector index at '${file}' cannot be used: ${detail} ` +
+            `An unreadable index and an empty one are different facts, and only one of them ` +
+            `is safe to answer with "no matches" — so this refuses rather than quietly ` +
+            `answering every question from the model alone on top of a file that already ` +
+            `exists. ` +
+            (problem === 'newer-schema'
+                ? `An index written by a newer agentfootprint needs a runtime that knows that ` +
+                    `schema; roll forward, or point this one at its own file.`
+                : problem === 'not-our-schema'
+                    ? `Give the store a file of its own rather than sharing one with tables of ` +
+                        `the same name.`
+                    : `Check the path and its permissions, or move the file aside to start a ` + `new one.`));
+        this.name = 'UnreadableIndexFileError';
+        this.file = file;
+        this.problem = problem;
+    }
+}
+exports.UnreadableIndexFileError = UnreadableIndexFileError;
+/**
+ * Raised when a vector meets an index built by a different embedder.
+ *
+ * **This is the refusal that keeps the store honest.** Cosine similarity between
+ * two different embedding spaces is not a weak signal — it is not a signal at
+ * all, and it comes back as a confident number in the same 0-to-1 range as a
+ * real one. There is no threshold that separates them, and nothing downstream
+ * can tell them apart. So the mismatch is refused where it happens, on both
+ * sides:
+ *
+ *  - at **write**, so a second embedder's vectors never enter a namespace;
+ *  - at **query**, so a swapped embedder never scores against the old ones.
+ *
+ * The named fix is an explicit re-index — delete the namespace and build it
+ * again with one embedder, or point the retriever at a different file. It is
+ * never a fallback: silently ignoring the mismatch is the failure, and silently
+ * re-embedding somebody's corpus is a bill they did not agree to.
+ *
+ * `problem` says which half is wrong. `'dimensions'` is arithmetically
+ * impossible to score at all; `'model'` would score, and lie.
+ */
+class EmbedderMismatchError extends Error {
+    code = 'ERR_EMBEDDER_MISMATCH';
+    /** The fingerprint the namespace was built with, `'<id>@<dims>'`. */
+    indexed;
+    /** The fingerprint that just arrived. */
+    incoming;
+    /** Which half disagrees. */
+    problem;
+    constructor(namespace, indexed, incoming, problem, operation) {
+        super(`[memory] cannot ${operation} the namespace '${namespace}': it was indexed by ` +
+            `'${indexed}' and this vector is from '${incoming}'. ` +
+            (problem === 'dimensions'
+                ? `Vectors of different lengths cannot be compared at all. `
+                : `Cosine similarity between two embedding spaces is not a weak signal — it is ` +
+                    `not a signal, and it comes back as a confident number in the same range as a ` +
+                    `real one, which no threshold can separate. `) +
+            `Re-index this namespace with one embedder (delete it and build it again), or ` +
+            `point this store at a different file. This refuses rather than re-embedding ` +
+            `your corpus on your behalf — that is a bill you did not agree to — and rather ` +
+            `than mixing the two, which would silently corrupt every ranking it touched.`);
+        this.name = 'EmbedderMismatchError';
+        this.indexed = indexed;
+        this.incoming = incoming;
+        this.problem = problem;
+    }
+}
+exports.EmbedderMismatchError = EmbedderMismatchError;
+// ─── The schema ──────────────────────────────────────────────────────
+/**
+ * What this runtime writes. Bumped only for a change an older reader could not
+ * survive — and an older reader meeting a newer number refuses by name rather
+ * than reading what it half-understands.
+ */
+const SCHEMA_VERSION = 1;
+const VECTORS_TABLE = 'af_vectors';
+const SIGNATURES_TABLE = 'af_signatures';
+const FEEDBACK_TABLE = 'af_feedback';
+const META_TABLE = 'af_index_meta';
+/**
+ * `dims`, `embedder_fp` and `source_uri` are COLUMNS as well as facts inside the
+ * JSON, and the redundancy is deliberate: it lets the file answer "what is in
+ * this index, and who built it?" from the `sqlite3` command line during an
+ * incident, without a JSON parser and without this library. An index you cannot
+ * inspect with the tools already on the box is one you debug by guessing.
+ */
+const VECTORS_COLUMNS = [
+    'namespace',
+    'id',
+    'value',
+    'metadata',
+    'embedding',
+    'dims',
+    'embedder_fp',
+    'version',
+    'created_at',
+    'updated_at',
+    'last_accessed_at',
+    'access_count',
+    'ttl',
+    'tier',
+    'source',
+    'source_uri',
+];
+/**
+ * Open (or create) a vector index in one SQLite file.
+ *
+ * @throws SqliteUnavailableError when the running Node has no `node:sqlite`.
+ * @throws UnreadableIndexFileError when the file exists but cannot be used —
+ *   never answered with an empty index.
+ * @throws EmbedderMismatchError from `put`/`putMany`/`search` when a vector
+ *   meets a namespace built by a different embedder.
+ *
+ * @example  Embed the corpus once, ever
+ * ```ts
+ * import { indexDocuments, defineRAG } from 'agentfootprint';
+ * import { sqliteVectorStore } from 'agentfootprint/memory';
+ * import { staticEmbedder } from 'agentfootprint/providers';
+ *
+ * const store = sqliteVectorStore({ file: './corpus.db' });
+ * const embedder = staticEmbedder();
+ *
+ * // First boot indexes; every boot after this one finds the vectors already there.
+ * await indexDocuments(store, embedder, docs, { embedderId: embedder.id });
+ *
+ * const agent = Agent.create({ provider })
+ *   .rag(defineRAG({ id: 'docs', store, embedder, embedderId: embedder.id }))
+ *   .build();
+ * ```
+ */
+function sqliteVectorStore(options) {
+    const { file, busyTimeoutMs = 5000 } = options;
+    if (file === ':memory:' || file.trim() === '') {
+        throw new TypeError(`[memory] sqliteVectorStore({ file: '${file}' }) is not a durable index. ` +
+            `':memory:' looks like a file and keeps nothing across a restart, so every boot ` +
+            `would re-embed the whole corpus — the one cost this store exists to remove. ` +
+            `Use InMemoryStore when an in-process index is what you want — it says so in ` +
+            `its name — or give this one a real path.`);
+    }
+    const DatabaseSync = options._sqlite ?? loadSqlite();
+    ensureParentDirectory(file);
+    let db;
+    try {
+        db = new DatabaseSync(file);
+    }
+    catch (err) {
+        throw new UnreadableIndexFileError(file, 'cannot-open', `${describe(err)}.`);
+    }
+    let journalMode;
+    try {
+        journalMode = applyPragmas(db, busyTimeoutMs);
+        ensureSchema(db, file);
+    }
+    catch (err) {
+        db.close();
+        if (err instanceof UnreadableIndexFileError)
+            throw err;
+        throw new UnreadableIndexFileError(file, 'cannot-open', `${describe(err)}.`);
+    }
+    const stmt = prepareStatements(db);
+    /** Resident matrices, one per namespace, dropped on any write to that namespace. */
+    const warm = new Map();
+    /** Fingerprints, read once per namespace and then held. */
+    const fingerprints = new Map();
+    let closed = false;
+    const open = (verb) => {
+        if (closed) {
+            throw new Error(`[memory] the sqliteVectorStore at '${file}' is closed, so it cannot ${verb}. ` +
+                `close() is final by design — reopening the file behind you would hide a ` +
+                `shutdown-ordering bug rather than surface it. Build a new store if you need ` +
+                `one after closing this.`);
+        }
+    };
+    const readFingerprint = (ns) => {
+        const held = fingerprints.get(ns);
+        if (held !== undefined)
+            return held;
+        const row = stmt.readFp.get(`fp:${ns}`);
+        if (typeof row?.value !== 'string')
+            return undefined;
+        fingerprints.set(ns, row.value);
+        return row.value;
+    };
+    const recordFingerprint = (ns, fp) => {
+        stmt.writeFp.run(`fp:${ns}`, fp);
+        fingerprints.set(ns, fp);
+    };
+    /**
+     * Compare an arriving fingerprint against the namespace's, refuse a real
+     * conflict, and adopt the arriving one when it is the first or the more
+     * specific of the two.
+     */
+    const reconcileFingerprint = (ns, incoming, operation) => {
+        const storedText = readFingerprint(ns);
+        if (storedText === undefined) {
+            if (operation === 'write to')
+                recordFingerprint(ns, fingerprintText(incoming));
+            return;
+        }
+        const stored = parseFingerprint(storedText);
+        const conflict = fingerprintConflict(stored, incoming);
+        if (conflict !== null) {
+            throw new EmbedderMismatchError(ns, storedText, fingerprintText(incoming), conflict, operation);
+        }
+        // Compatible. If the index recorded an anonymous embedder and this vector
+        // names one, upgrade the record — a named fingerprint can refuse a future
+        // swap that an anonymous one has to let through.
+        if (operation === 'write to' && stored.id === undefined && incoming.id !== undefined) {
+            recordFingerprint(ns, fingerprintText(incoming));
+        }
+    };
+    const writeEntry = (ns, entry) => {
+        const vector = toFloat32(entry.embedding);
+        if (vector) {
+            reconcileFingerprint(ns, {
+                ...(entry.embeddingModel !== undefined && { id: entry.embeddingModel }),
+                dims: vector.length,
+            }, 'write to');
+        }
+        stmt.upsert.run(ns, entry.id, JSON.stringify(entry.value ?? null), entry.metadata === undefined ? null : JSON.stringify(entry.metadata), vector ? blobOf(vector) : null, vector ? vector.length : 0, vector
+            ? fingerprintText({
+                ...(entry.embeddingModel !== undefined && { id: entry.embeddingModel }),
+                dims: vector.length,
+            })
+            : null, entry.version, entry.createdAt, entry.updatedAt, entry.lastAccessedAt, entry.accessCount, entry.ttl ?? null, entry.tier ?? null, entry.source === undefined ? null : JSON.stringify(entry.source), readSourceUri(entry), entry.embeddingModel ?? null);
+        warm.delete(ns);
+    };
+    const hydrate = (ns) => {
+        const held = warm.get(ns);
+        if (held)
+            return held;
+        const rows = stmt.vectorsOf.all(ns);
+        const dims = rows.length > 0 ? rows[0]?.dims ?? 0 : 0;
+        const matrix = new Float32Array(rows.length * dims);
+        const ids = [];
+        const ttls = [];
+        const tiers = [];
+        const embedderIds = [];
+        const degenerate = new Set();
+        let row = 0;
+        for (const r of rows) {
+            // A namespace cannot hold two vector lengths — the fingerprint refusal
+            // above is what guarantees it — but a file edited by hand can, and a
+            // ragged matrix is worse than a skipped row.
+            if (r.dims !== dims)
+                continue;
+            const vec = readFloat32(r.embedding, dims);
+            let norm = 0;
+            for (let i = 0; i < dims; i++)
+                norm += (vec[i] ?? 0) * (vec[i] ?? 0);
+            norm = Math.sqrt(norm);
+            const offset = row * dims;
+            if (norm === 0) {
+                degenerate.add(row);
+            }
+            else {
+                // Normalised at hydration, so every query is a dot product. The BLOB on
+                // disk keeps the ORIGINAL vector, so `get`/`list` round-trip exactly
+                // what the caller wrote.
+                for (let i = 0; i < dims; i++)
+                    matrix[offset + i] = (vec[i] ?? 0) / norm;
+            }
+            ids.push(r.id);
+            ttls.push(r.ttl ?? undefined);
+            tiers.push(r.tier ?? undefined);
+            embedderIds.push(r.embedding_model ?? undefined);
+            row += 1;
+        }
+        const built = { ids, matrix, dims, ttls, tiers, embedderIds, degenerate };
+        warm.set(ns, built);
+        return built;
+    };
+    const store = {
+        journalMode,
+        file,
+        fingerprintOf(identity) {
+            open('report a fingerprint');
+            return readFingerprint((0, index_js_1.identityNamespace)(identity));
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async warm(identity) {
+            open('warm a namespace');
+            const startedAt = Date.now();
+            const warmed = hydrate((0, index_js_1.identityNamespace)(identity));
+            return { count: warmed.ids.length, durationMs: Date.now() - startedAt };
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async get(identity, id) {
+            open('read an entry');
+            const ns = (0, index_js_1.identityNamespace)(identity);
+            const row = stmt.select.get(ns, id);
+            if (row === undefined)
+                return null;
+            if (isExpired(row.ttl))
+                return null;
+            // Decay signals, the same side effect the port documents for `get`. It
+            // does NOT invalidate the warm matrix: access counters are not vectors.
+            stmt.touch.run(Date.now(), ns, id);
+            return rowToEntry(row);
+        },
+        async put(identity, entry) {
+            await this.putMany(identity, [entry]);
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async putMany(identity, entries) {
+            open('write entries');
+            // The port requires an empty batch to be a no-op — callers rely on it to
+            // skip a round-trip on a turn that produced nothing.
+            if (entries.length === 0)
+                return;
+            const ns = (0, index_js_1.identityNamespace)(identity);
+            // ONE transaction for the batch. The port says atomicity is not
+            // guaranteed across a batch and most callers are append-idempotent — but
+            // a HALF-INDEXED corpus is a specific kind of bad: retrieval keeps
+            // working and quietly cannot see the documents that did not land, which
+            // reads as "the model does not know that" rather than as a failure. A
+            // file can offer all-or-nothing cheaply, so it does.
+            db.exec('BEGIN');
+            try {
+                for (const entry of entries)
+                    writeEntry(ns, entry);
+                db.exec('COMMIT');
+            }
+            catch (err) {
+                try {
+                    db.exec('ROLLBACK');
+                }
+                catch {
+                    /* The original failure is the one worth reporting. */
+                }
+                warm.delete(ns);
+                throw err;
+            }
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async putIfVersion(identity, entry, expectedVersion) {
+            open('write an entry');
+            const ns = (0, index_js_1.identityNamespace)(identity);
+            // Read and write inside one transaction, so the check and the write
+            // cannot be separated by another writer — which is the entire point of a
+            // compare-and-set.
+            db.exec('BEGIN IMMEDIATE');
+            try {
+                const row = stmt.selectVersion.get(ns, entry.id);
+                const current = row?.version;
+                if (current === undefined) {
+                    if (expectedVersion !== 0) {
+                        db.exec('COMMIT');
+                        return { applied: false };
+                    }
+                }
+                else if (current !== expectedVersion) {
+                    db.exec('COMMIT');
+                    return { applied: false, currentVersion: current };
+                }
+                writeEntry(ns, entry);
+                db.exec('COMMIT');
+                return { applied: true };
+            }
+            catch (err) {
+                try {
+                    db.exec('ROLLBACK');
+                }
+                catch {
+                    /* The original failure is the one worth reporting. */
+                }
+                warm.delete(ns);
+                throw err;
+            }
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async list(identity, listOptions) {
+            open('list entries');
+            const ns = (0, index_js_1.identityNamespace)(identity);
+            const limit = Math.max(1, Math.floor(listOptions?.limit ?? 100));
+            const after = listOptions?.cursor ?? '';
+            const rows = stmt.list.all(ns, after, limit + 1);
+            const tierFilter = listOptions?.tiers ? new Set(listOptions.tiers) : undefined;
+            const page = [];
+            let cursor;
+            for (const row of rows) {
+                if (page.length === limit) {
+                    // The extra row only ever exists to prove there IS a next page.
+                    cursor = page[page.length - 1]?.id;
+                    break;
+                }
+                if (isExpired(row.ttl))
+                    continue;
+                if (tierFilter && (row.tier === null || !tierFilter.has(row.tier)))
+                    continue;
+                page.push(rowToEntry(row));
+            }
+            return { entries: page, ...(cursor !== undefined && { cursor }) };
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async delete(identity, id) {
+            open('delete an entry');
+            const ns = (0, index_js_1.identityNamespace)(identity);
+            stmt.remove.run(ns, id);
+            warm.delete(ns);
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async seen(identity, signature) {
+            open('check a signature');
+            return stmt.seen.get((0, index_js_1.identityNamespace)(identity), signature) !== undefined;
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async recordSignature(identity, signature) {
+            open('record a signature');
+            stmt.addSignature.run((0, index_js_1.identityNamespace)(identity), signature);
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async feedback(identity, id, usefulness) {
+            open('record feedback');
+            // Non-finite values poison the aggregate; the port says adapters must
+            // reject them, and clamp the rest.
+            if (!Number.isFinite(usefulness))
+                return;
+            const clamped = Math.max(-1, Math.min(1, usefulness));
+            stmt.addFeedback.run((0, index_js_1.identityNamespace)(identity), id, clamped);
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async getFeedback(identity, id) {
+            open('read feedback');
+            const row = stmt.readFeedback.get((0, index_js_1.identityNamespace)(identity), id);
+            const count = row?.count ?? 0;
+            if (count === 0)
+                return null;
+            return { average: (row?.total ?? 0) / count, count };
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async forget(identity) {
+            open('forget a namespace');
+            const ns = (0, index_js_1.identityNamespace)(identity);
+            db.exec('BEGIN');
+            try {
+                stmt.forgetVectors.run(ns);
+                stmt.forgetSignatures.run(ns);
+                stmt.forgetFeedback.run(ns);
+                stmt.forgetMeta.run(`fp:${ns}`);
+                db.exec('COMMIT');
+            }
+            catch (err) {
+                try {
+                    db.exec('ROLLBACK');
+                }
+                catch {
+                    /* The original failure is the one worth reporting. */
+                }
+                throw err;
+            }
+            warm.delete(ns);
+            fingerprints.delete(ns);
+        },
+        // eslint-disable-next-line @typescript-eslint/require-await
+        async search(identity, query, searchOptions) {
+            open('search');
+            const ns = (0, index_js_1.identityNamespace)(identity);
+            const k = Math.max(1, Math.floor(searchOptions?.k ?? 10));
+            // Refused BEFORE the scan, on both halves of the fingerprint. A swapped
+            // embedder must not score against the old vectors — the numbers come back
+            // in the same range as real ones and no threshold separates them.
+            reconcileFingerprint(ns, {
+                ...(searchOptions?.embedderId !== undefined && { id: searchOptions.embedderId }),
+                dims: query.length,
+            }, 'search');
+            const warmed = hydrate(ns);
+            if (warmed.ids.length === 0 || warmed.dims === 0)
+                return [];
+            if (warmed.dims !== query.length)
+                return [];
+            // Normalise the query once; the matrix rows already are. Cosine is then
+            // one dot product per row and nothing else.
+            let qNorm = 0;
+            for (const v of query)
+                qNorm += v * v;
+            qNorm = Math.sqrt(qNorm);
+            if (qNorm === 0)
+                return [];
+            const q = new Float32Array(query.length);
+            for (let i = 0; i < query.length; i++)
+                q[i] = (query[i] ?? 0) / qNorm;
+            const now = Date.now();
+            const tierFilter = searchOptions?.tiers ? new Set(searchOptions.tiers) : undefined;
+            const embedderId = searchOptions?.embedderId;
+            const minScore = searchOptions?.minScore;
+            const { dims, matrix, ids } = warmed;
+            const scored = [];
+            for (let row = 0; row < ids.length; row++) {
+                if (warmed.degenerate.has(row))
+                    continue;
+                const ttl = warmed.ttls[row];
+                if (ttl !== undefined && ttl <= now)
+                    continue;
+                if (tierFilter) {
+                    const tier = warmed.tiers[row];
+                    if (tier === undefined || !tierFilter.has(tier))
+                        continue;
+                }
+                if (embedderId !== undefined) {
+                    const rowEmbedder = warmed.embedderIds[row];
+                    if (rowEmbedder !== undefined && rowEmbedder !== embedderId)
+                        continue;
+                }
+                const offset = row * dims;
+                let dot = 0;
+                for (let i = 0; i < dims; i++)
+                    dot += (matrix[offset + i] ?? 0) * (q[i] ?? 0);
+                if (minScore !== undefined && dot < minScore)
+                    continue;
+                // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+                scored.push({ id: ids[row], score: dot });
+            }
+            scored.sort((a, b) => {
+                if (b.score !== a.score)
+                    return b.score - a.score;
+                return a.id < b.id ? -1 : a.id > b.id ? 1 : 0;
+            });
+            // Hydrate only the winners' payloads. The matrix holds vectors, not
+            // documents — a 50,000-chunk namespace has no business materialising
+            // 50,000 JSON values to answer a top-3.
+            const out = [];
+            for (const hit of scored.slice(0, k)) {
+                const row = stmt.select.get(ns, hit.id);
+                if (row === undefined)
+                    continue;
+                out.push({ entry: rowToEntry(row), score: hit.score });
+            }
+            return out;
+        },
+        close() {
+            if (closed)
+                return;
+            closed = true;
+            warm.clear();
+            db.close();
+        },
+    };
+    return store;
+}
+exports.sqliteVectorStore = sqliteVectorStore;
+/** `'<id>@<dims>'`, with `'?'` for an embedder that did not name itself. */
+function fingerprintText(fp) {
+    return `${fp.id ?? '?'}@${fp.dims}`;
+}
+function parseFingerprint(text) {
+    const at = text.lastIndexOf('@');
+    const id = at === -1 ? '?' : text.slice(0, at);
+    const dims = at === -1 ? 0 : Number(text.slice(at + 1));
+    return {
+        ...(id !== '?' && id !== '' && { id }),
+        dims: Number.isFinite(dims) ? dims : 0,
+    };
+}
+/**
+ * What, if anything, makes these two incompatible.
+ *
+ * Dimensions always decide: two lengths cannot be compared at all. Model ids
+ * decide only when BOTH sides named themselves — an anonymous vector is not
+ * evidence of a different embedder, and refusing on absence would break every
+ * caller who never passed an `embedderId`, which is most of them.
+ */
+function fingerprintConflict(stored, incoming) {
+    if (stored.dims !== incoming.dims)
+        return 'dimensions';
+    if (stored.id !== undefined && incoming.id !== undefined && stored.id !== incoming.id) {
+        return 'model';
+    }
+    return null;
+}
+function rowToEntry(row) {
+    return {
+        id: row.id,
+        value: parseJson(row.value),
+        ...(row.metadata !== null && {
+            metadata: parseJson(row.metadata),
+        }),
+        version: row.version,
+        createdAt: row.created_at,
+        updatedAt: row.updated_at,
+        lastAccessedAt: row.last_accessed_at,
+        accessCount: row.access_count,
+        ...(row.ttl !== null && { ttl: row.ttl }),
+        ...(row.tier !== null && { tier: row.tier }),
+        ...(row.source !== null && {
+            source: parseJson(row.source),
+        }),
+        ...(row.embedding !== null &&
+            row.dims > 0 && { embedding: Array.from(readFloat32(row.embedding, row.dims)) }),
+        ...(row.embedding_model !== null && { embeddingModel: row.embedding_model }),
+    };
+}
+function parseJson(text) {
+    try {
+        return JSON.parse(text);
+    }
+    catch {
+        // A value this store could not read back. It was written as JSON by this
+        // store, so reaching here means the file was edited by something else.
+        return null;
+    }
+}
+function isExpired(ttl) {
+    return ttl !== null && ttl <= Date.now();
+}
+/** The document a chunk came from, lifted to a column so `sqlite3` can group by it. */
+function readSourceUri(entry) {
+    const value = entry.value;
+    const meta = (value?.metadata ?? entry.metadata);
+    const uri = meta?.['docUri'] ?? meta?.['source'];
+    return typeof uri === 'string' && uri.length > 0 ? uri : null;
+}
+// ─── Vector encoding ─────────────────────────────────────────────────
+function toFloat32(embedding) {
+    if (!embedding || embedding.length === 0)
+        return undefined;
+    const out = new Float32Array(embedding.length);
+    for (let i = 0; i < embedding.length; i++)
+        out[i] = embedding[i] ?? 0;
+    return out;
+}
+/** Little-endian `Float32Array` bytes — the platform's own layout, no conversion. */
+function blobOf(vector) {
+    return new Uint8Array(vector.buffer, vector.byteOffset, vector.byteLength);
+}
+function readFloat32(blob, dims) {
+    // A BLOB comes back with its own byteOffset; a Float32Array view needs a
+    // 4-byte-aligned one, which is not guaranteed, so copy when it is not.
+    if (blob.byteOffset % 4 === 0) {
+        return new Float32Array(blob.buffer, blob.byteOffset, Math.min(dims, blob.byteLength / 4));
+    }
+    const copy = new Uint8Array(blob);
+    return new Float32Array(copy.buffer, 0, Math.min(dims, copy.byteLength / 4));
+}
+// ─── Internals ───────────────────────────────────────────────────────
+function loadSqlite() {
+    try {
+        const mod = (0, lazyRequire_js_1.lazyRequire)('node:sqlite');
+        if (typeof mod?.DatabaseSync !== 'function') {
+            throw new Error('the module loaded but has no DatabaseSync');
+        }
+        return mod.DatabaseSync;
+    }
+    catch (err) {
+        throw new sqliteUnavailable_js_1.SqliteUnavailableError(nodeVersion(), describe(err), {
+            door: 'memory',
+            factory: 'sqliteVectorStore()',
+            alternative: 'InMemoryStore — which keeps the index in a Map and re-embeds the whole corpus on every restart, and says so in its name',
+            whyNotFallback: 'an index that silently forgot every document on restart looks, from the outside, exactly like a corpus that was never built',
+        });
+    }
+}
+function nodeVersion() {
+    return typeof process !== 'undefined' && typeof process.version === 'string'
+        ? process.version
+        : 'unknown';
+}
+function ensureParentDirectory(file) {
+    try {
+        const { mkdirSync } = (0, lazyRequire_js_1.lazyRequire)('node:fs');
+        const { dirname } = (0, lazyRequire_js_1.lazyRequire)('node:path');
+        mkdirSync(dirname(file), { recursive: true });
+    }
+    catch {
+        /* Reported by the open below, with the path and the real reason. */
+    }
+}
+/**
+ * Set the file up for concurrent use and report what it actually got.
+ *
+ * `journal_mode` runs first on purpose: it is the point at which SQLite reads
+ * the file header, so "this is not a database" surfaces here, at construction,
+ * rather than on the first query of the night.
+ */
+function applyPragmas(db, busyTimeoutMs) {
+    const row = db.prepare('PRAGMA journal_mode = WAL').get();
+    // A pragma takes a literal, not a bound parameter, so the number is coerced
+    // to one before it reaches the statement rather than trusted to be one.
+    const waitMs = Number.isFinite(busyTimeoutMs) ? Math.max(0, Math.floor(busyTimeoutMs)) : 5000;
+    db.exec(`PRAGMA busy_timeout = ${waitMs}`);
+    // NORMAL, not FULL: with WAL that is durable across a process crash — which
+    // is what this store promises — without a disk sync per write. A power cut
+    // can still cost the most recent commits; the answer to that is a
+    // fleet-grade store, not a pragma.
+    db.exec('PRAGMA synchronous = NORMAL');
+    return typeof row?.journal_mode === 'string' ? row.journal_mode : 'unknown';
+}
+function ensureSchema(db, file) {
+    db.exec(`CREATE TABLE IF NOT EXISTS ${VECTORS_TABLE} (` +
+        `namespace TEXT NOT NULL, id TEXT NOT NULL, value TEXT NOT NULL, metadata TEXT, ` +
+        `embedding BLOB, dims INTEGER NOT NULL, embedder_fp TEXT, version INTEGER NOT NULL, ` +
+        `created_at INTEGER NOT NULL, updated_at INTEGER NOT NULL, ` +
+        `last_accessed_at INTEGER NOT NULL, access_count INTEGER NOT NULL, ` +
+        `ttl INTEGER, tier TEXT, source TEXT, source_uri TEXT, embedding_model TEXT, ` +
+        `PRIMARY KEY (namespace, id)) STRICT`);
+    // The identity check runs HERE — after the table, before anything that
+    // depends on its columns. `CREATE TABLE IF NOT EXISTS` succeeds against
+    // somebody else's table of the same name, and the very next statement
+    // (an index over `namespace`) would then fail on a missing column and be
+    // reported as 'cannot-open' — the right refusal for the wrong reason,
+    // telling the reader to check permissions on a file whose real problem is
+    // that it belongs to something else.
+    const columns = db.prepare(`PRAGMA table_info('${VECTORS_TABLE}')`).all().map((c) => String(c.name));
+    const missing = VECTORS_COLUMNS.filter((name) => !columns.includes(name));
+    if (missing.length > 0) {
+        throw new UnreadableIndexFileError(file, 'not-our-schema', `it has an '${VECTORS_TABLE}' table that is not this store's ` +
+            `(missing ${missing.join(', ')}; found ${columns.join(', ') || 'nothing'}).`);
+    }
+    db.exec(`CREATE INDEX IF NOT EXISTS ${VECTORS_TABLE}_ns ON ${VECTORS_TABLE}(namespace)`);
+    db.exec(`CREATE INDEX IF NOT EXISTS ${VECTORS_TABLE}_doc ON ${VECTORS_TABLE}(namespace, source_uri)`);
+    db.exec(`CREATE TABLE IF NOT EXISTS ${SIGNATURES_TABLE} (` +
+        `namespace TEXT NOT NULL, signature TEXT NOT NULL, PRIMARY KEY (namespace, signature)) STRICT`);
+    db.exec(`CREATE TABLE IF NOT EXISTS ${FEEDBACK_TABLE} (` +
+        `namespace TEXT NOT NULL, id TEXT NOT NULL, total REAL NOT NULL, count INTEGER NOT NULL, ` +
+        `PRIMARY KEY (namespace, id)) STRICT`);
+    db.exec(`CREATE TABLE IF NOT EXISTS ${META_TABLE} (key TEXT PRIMARY KEY, value TEXT NOT NULL) STRICT`);
+    const stored = db
+        .prepare(`SELECT value FROM ${META_TABLE} WHERE key = 'schema_version'`)
+        .get();
+    if (stored === undefined) {
+        db.prepare(`INSERT INTO ${META_TABLE} (key, value) VALUES ('schema_version', ?)`).run(String(SCHEMA_VERSION));
+        return;
+    }
+    const found = Number(stored.value);
+    if (!Number.isFinite(found) || found > SCHEMA_VERSION) {
+        throw new UnreadableIndexFileError(file, 'newer-schema', `it was written with schema version ${String(stored.value)} and this runtime ` +
+            `reads version ${SCHEMA_VERSION}.`);
+    }
+}
+function prepareStatements(db) {
+    const ENTRY_COLUMNS = 'id, value, metadata, embedding, dims, version, created_at, updated_at, ' +
+        'last_accessed_at, access_count, ttl, tier, source, embedding_model';
+    return {
+        upsert: db.prepare(`INSERT INTO ${VECTORS_TABLE} (namespace, id, value, metadata, embedding, dims, ` +
+            `embedder_fp, version, created_at, updated_at, last_accessed_at, access_count, ` +
+            `ttl, tier, source, source_uri, embedding_model) ` +
+            `VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?) ` +
+            `ON CONFLICT(namespace, id) DO UPDATE SET value = excluded.value, ` +
+            `metadata = excluded.metadata, embedding = excluded.embedding, dims = excluded.dims, ` +
+            `embedder_fp = excluded.embedder_fp, version = excluded.version, ` +
+            `created_at = excluded.created_at, updated_at = excluded.updated_at, ` +
+            `last_accessed_at = excluded.last_accessed_at, access_count = excluded.access_count, ` +
+            `ttl = excluded.ttl, tier = excluded.tier, source = excluded.source, ` +
+            `source_uri = excluded.source_uri, embedding_model = excluded.embedding_model`),
+        select: db.prepare(`SELECT ${ENTRY_COLUMNS} FROM ${VECTORS_TABLE} WHERE namespace = ? AND id = ?`),
+        selectVersion: db.prepare(`SELECT version FROM ${VECTORS_TABLE} WHERE namespace = ? AND id = ?`),
+        list: db.prepare(`SELECT ${ENTRY_COLUMNS} FROM ${VECTORS_TABLE} WHERE namespace = ? AND id > ? ` +
+            `ORDER BY id LIMIT ?`),
+        vectorsOf: db.prepare(`SELECT id, embedding, dims, ttl, tier, embedding_model FROM ${VECTORS_TABLE} ` +
+            `WHERE namespace = ? AND embedding IS NOT NULL ORDER BY id`),
+        touch: db.prepare(`UPDATE ${VECTORS_TABLE} SET last_accessed_at = ?, access_count = access_count + 1 ` +
+            `WHERE namespace = ? AND id = ?`),
+        remove: db.prepare(`DELETE FROM ${VECTORS_TABLE} WHERE namespace = ? AND id = ?`),
+        seen: db.prepare(`SELECT 1 FROM ${SIGNATURES_TABLE} WHERE namespace = ? AND signature = ?`),
+        addSignature: db.prepare(`INSERT INTO ${SIGNATURES_TABLE} (namespace, signature) VALUES (?, ?) ` +
+            `ON CONFLICT(namespace, signature) DO NOTHING`),
+        addFeedback: db.prepare(`INSERT INTO ${FEEDBACK_TABLE} (namespace, id, total, count) VALUES (?, ?, ?, 1) ` +
+            `ON CONFLICT(namespace, id) DO UPDATE SET total = total + excluded.total, ` +
+            `count = count + 1`),
+        readFeedback: db.prepare(`SELECT total, count FROM ${FEEDBACK_TABLE} WHERE namespace = ? AND id = ?`),
+        forgetVectors: db.prepare(`DELETE FROM ${VECTORS_TABLE} WHERE namespace = ?`),
+        forgetSignatures: db.prepare(`DELETE FROM ${SIGNATURES_TABLE} WHERE namespace = ?`),
+        forgetFeedback: db.prepare(`DELETE FROM ${FEEDBACK_TABLE} WHERE namespace = ?`),
+        forgetMeta: db.prepare(`DELETE FROM ${META_TABLE} WHERE key = ?`),
+        readFp: db.prepare(`SELECT value FROM ${META_TABLE} WHERE key = ?`),
+        writeFp: db.prepare(`INSERT INTO ${META_TABLE} (key, value) VALUES (?, ?) ` +
+            `ON CONFLICT(key) DO UPDATE SET value = excluded.value`),
+    };
+}
+/** One sentence about a thrown thing, for a message that has to stay readable. */
+function describe(err) {
+    return err instanceof Error ? err.message : String(err);
+}
+//# sourceMappingURL=sqliteVector.js.map