agentdb 3.1.0 → 3.2.0

package/dist/src/search/cross-encoder.js

@@ -0,0 +1,52 @@
+/**
+ * Opt-in cross-encoder reranking (ADR-016). A cross-encoder reads the query and a candidate together, so it ranks
+ * far better than either embedding similarity or BM25, at about 200 ms per 60 candidates on CPU for the default
+ * 22M-parameter model (q8). It is therefore applied only to a bounded candidate set, and only when a caller asks.
+ *
+ * The model is downloaded once by @huggingface/transformers. Every input is bounded: at most MAX_CANDIDATES
+ * candidates, each cut to MAX_CHARS before tokenising, and the tokenizer truncates pairs at 512 tokens.
+ */
+export const DEFAULT_RERANKER = process.env.AGENTDB_RERANKER_MODEL || 'Xenova/ms-marco-MiniLM-L-6-v2';
+/** Two candidate lists of RERANK_DEPTH (60) each, deduplicated, fit. */
+export const MAX_CANDIDATES = 128;
+const MAX_CHARS = 4_096;
+const BATCH = 32;
+const loading = new Map();
+async function load(modelId) {
+    let p = loading.get(modelId);
+    if (!p) {
+        p = (async () => {
+            const t = await import('@huggingface/transformers');
+            const tokenizer = await t.AutoTokenizer.from_pretrained(modelId);
+            const model = await t.AutoModelForSequenceClassification.from_pretrained(modelId, { dtype: 'q8' });
+            return { tokenizer: tokenizer, model: model };
+        })();
+        // A failed load is not cached, so a later call can retry (e.g. after the network returns).
+        p.catch(() => loading.delete(modelId));
+        loading.set(modelId, p);
+    }
+    return p;
+}
+/**
+ * Relevance logits of each candidate for the query, in input order. Throws if the model cannot be loaded;
+ * callers fall back to their unreranked order.
+ */
+export async function crossEncoderScores(query, candidates, modelId = DEFAULT_RERANKER) {
+    if (candidates.length > MAX_CANDIDATES)
+        throw new RangeError(`at most ${MAX_CANDIDATES} candidates`);
+    const out = new Float64Array(candidates.length);
+    if (candidates.length === 0)
+        return out;
+    const { tokenizer, model } = await load(modelId);
+    const q = query.slice(0, MAX_CHARS);
+    for (let i = 0; i < candidates.length; i += BATCH) {
+        const docs = candidates.slice(i, i + BATCH).map(d => d.slice(0, MAX_CHARS));
+        const enc = tokenizer(new Array(docs.length).fill(q), { text_pair: docs, padding: true, truncation: true, max_length: 512 });
+        const { logits } = await model(enc);
+        const width = logits.dims[1] ?? 1;
+        for (let j = 0; j < docs.length; j++)
+            out[i + j] = Number(logits.data[j * width + (width - 1)]);
+    }
+    return out;
+}
+//# sourceMappingURL=cross-encoder.js.map

package/dist/src/search/cross-encoder.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cross-encoder.js","sourceRoot":"","sources":["../../../src/search/cross-encoder.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,CAAC,MAAM,gBAAgB,GAAG,OAAO,CAAC,GAAG,CAAC,sBAAsB,IAAI,+BAA+B,CAAC;AACtG,wEAAwE;AACxE,MAAM,CAAC,MAAM,cAAc,GAAG,GAAG,CAAC;AAClC,MAAM,SAAS,GAAG,KAAK,CAAC;AACxB,MAAM,KAAK,GAAG,EAAE,CAAC;AAOjB,MAAM,OAAO,GAAG,IAAI,GAAG,EAA2B,CAAC;AAEnD,KAAK,UAAU,IAAI,CAAC,OAAe;IACjC,IAAI,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;IAC7B,IAAI,CAAC,CAAC,EAAE,CAAC;QACP,CAAC,GAAG,CAAC,KAAK,IAAI,EAAE;YACd,MAAM,CAAC,GAAG,MAAM,MAAM,CAAC,2BAA2B,CAAC,CAAC;YACpD,MAAM,SAAS,GAAG,MAAM,CAAC,CAAC,aAAa,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC;YACjE,MAAM,KAAK,GAAG,MAAM,CAAC,CAAC,kCAAkC,CAAC,eAAe,CAAC,OAAO,EAAE,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;YACnG,OAAO,EAAE,SAAS,EAAE,SAA2C,EAAE,KAAK,EAAE,KAAmC,EAAE,CAAC;QAChH,CAAC,CAAC,EAAE,CAAC;QACL,2FAA2F;QAC3F,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,OAAO,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC;QACvC,OAAO,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;IAC1B,CAAC;IACD,OAAO,CAAC,CAAC;AACX,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAC,KAAa,EAAE,UAA6B,EAAE,OAAO,GAAG,gBAAgB;IAC/G,IAAI,UAAU,CAAC,MAAM,GAAG,cAAc;QAAE,MAAM,IAAI,UAAU,CAAC,WAAW,cAAc,aAAa,CAAC,CAAC;IACrG,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;IAChD,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,GAAG,CAAC;IACxC,MAAM,EAAE,SAAS,EAAE,KAAK,EAAE,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,CAAC;IACjD,MAAM,CAAC,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC;IACpC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,UAAU,CAAC,MAAM,EAAE,CAAC,IAAI,KAAK,EAAE,CAAC;QAClD,MAAM,IAAI,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC,CAAC;QAC5E,MAAM,GAAG,GAAG,SAAS,CAAC,IAAI,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,UAAU,EAAE,IAAI,EAAE,UAAU,EAAE,GAAG,EAAE,CAAC,CAAC;QAC7H,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,KAAK,CAAC,GAAG,CAAC,CAAC;QACpC,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QAClC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE;YAAE,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IAClG,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentdb",
-  "version": "3.1.0",
+  "version": "3.2.0",
   "description": "Self-learning vector memory for AI agents — single-file .rvf cognitive container with HNSW search, episodic Reflexion memory, causal graph + Cypher, 9 RL algorithms, Thompson Sampling bandit, 41 MCP tools, hybrid (BM25 + dense) retrieval, GNN attention. 150x faster than SQLite. Up to +36% search quality from feedback. Runs in Node, browsers (WASM), edge, offline. Backs the agentic-flow / ruflo / Claude Code memory layer.",
   "type": "module",
   "main": "dist/src/index.js",
@@ -210,7 +210,7 @@
     "commander": "^12.1.0",
     "hnswlib-node": "^3.0.0",
     "inquirer": "^9.3.8",
-    "ruvector": "^0.2.40",
+    "ruvector": "^0.3.5",
     "ruvector-attention-wasm": "^0.1.32",
     "ruvector-graph-transformer-wasm": "^2.0.4"
   },