{"version":3,"file":"reranker-HHmhZk6v.mjs","names":[],"sources":["../src/memory/reranker.ts"],"sourcesContent":["/**\n * Cross-encoder reranker for PAI memory search results.\n *\n * Uses Xenova/ms-marco-MiniLM-L-6-v2 — a 22.7M param cross-encoder trained on\n * MS MARCO passage ranking. The q8 quantized ONNX model is ~23 MB.\n *\n * Cross-encoders score (query, document) pairs jointly, producing more accurate\n * relevance scores than bi-encoder cosine similarity alone. The trade-off is\n * latency: cross-encoders must score each pair independently, so they are used\n * as a reranking step on top of a fast first-stage retriever (BM25 / cosine).\n *\n * The model is loaded as a lazy singleton — no startup cost until the first\n * rerank call. Subsequent calls reuse the loaded model.\n *\n * Inspired by QMD's Qwen3-reranker step (tobi/qmd).\n */\n\nimport type { SearchResult } from \"./search.js\";\n\n// ---------------------------------------------------------------------------\n// Constants\n// ---------------------------------------------------------------------------\n\nconst DEFAULT_RERANKER_MODEL = \"Xenova/ms-marco-MiniLM-L-6-v2\";\n\n// ---------------------------------------------------------------------------\n// Lazy singleton\n// ---------------------------------------------------------------------------\n\nlet _tokenizer: any = null;\nlet _model: any = null;\nlet _currentModel: string | null = null;\nlet _loading: Promise | null = null;\n\n/**\n * Configure the reranker model.\n * Must be called before the first rerank() call if you want a non-default model.\n */\nexport function configureRerankerModel(model?: string): void {\n const resolved = model?.trim() || DEFAULT_RERANKER_MODEL;\n if (_currentModel !== null && _currentModel !== resolved) {\n _tokenizer = null;\n _model = null;\n _loading = null;\n }\n _currentModel = resolved;\n}\n\nasync function ensureLoaded(): Promise {\n if (_tokenizer && _model) return;\n if (_loading) return _loading;\n\n _loading = (async () => {\n const model = _currentModel ?? DEFAULT_RERANKER_MODEL;\n const {\n AutoTokenizer,\n AutoModelForSequenceClassification,\n } = await import(\"@huggingface/transformers\");\n\n _tokenizer = await AutoTokenizer.from_pretrained(model);\n _model = await AutoModelForSequenceClassification.from_pretrained(\n model,\n { dtype: \"q8\" },\n );\n _currentModel = model;\n })();\n\n return _loading;\n}\n\n// ---------------------------------------------------------------------------\n// Reranking\n// ---------------------------------------------------------------------------\n\nexport interface RerankOptions {\n /** Maximum number of results to return after reranking. */\n topK?: number;\n /**\n * Maximum number of candidates to rerank.\n * Cross-encoders are O(n) per candidate, so we cap to keep latency\n * reasonable. Default: 50.\n */\n maxCandidates?: number;\n}\n\n/**\n * Rerank search results using a cross-encoder model.\n *\n * Takes the top `maxCandidates` results from a first-stage retriever,\n * scores each (query, snippet) pair through the cross-encoder, and\n * returns them sorted by cross-encoder relevance score.\n *\n * The original retrieval score is replaced with the cross-encoder score.\n */\nexport async function rerankResults(\n query: string,\n results: SearchResult[],\n opts?: RerankOptions,\n): Promise {\n if (results.length === 0) return [];\n\n const maxCandidates = opts?.maxCandidates ?? 50;\n const topK = opts?.topK ?? results.length;\n\n // Cap candidates to rerank\n const candidates = results.slice(0, maxCandidates);\n\n await ensureLoaded();\n\n // Tokenize all (query, document) pairs in a single batch\n const queries = new Array(candidates.length).fill(query);\n const documents = candidates.map((r) => r.snippet);\n\n const inputs = _tokenizer!(queries, {\n text_pair: documents,\n padding: true,\n truncation: true,\n });\n\n // Run the cross-encoder\n const output = await _model!(inputs);\n const logits = output.logits;\n\n // ms-marco-MiniLM returns raw logits (not sigmoid-normalized).\n // Higher = more relevant.\n const scores: number[][] = logits.tolist();\n\n // Build reranked results\n const scored = candidates.map((result, i) => ({\n ...result,\n score: scores[i][0],\n }));\n\n // Sort by cross-encoder score descending\n scored.sort((a, b) => b.score - a.score);\n\n return scored.slice(0, topK);\n}\n"],"mappings":";AAuBA,MAAM,yBAAyB;AAM/B,IAAI,aAAkB;AACtB,IAAI,SAAc;AAClB,IAAI,gBAA+B;AACnC,IAAI,WAAiC;;;;;AAMrC,SAAgB,uBAAuB,OAAsB;CAC3D,MAAM,WAAW,OAAO,MAAM,IAAI;AAClC,KAAI,kBAAkB,QAAQ,kBAAkB,UAAU;AACxD,eAAa;AACb,WAAS;AACT,aAAW;;AAEb,iBAAgB;;AAGlB,eAAe,eAA8B;AAC3C,KAAI,cAAc,OAAQ;AAC1B,KAAI,SAAU,QAAO;AAErB,aAAY,YAAY;EACtB,MAAM,QAAQ,iBAAiB;EAC/B,MAAM,EACJ,eACA,uCACE,MAAM,OAAO;AAEjB,eAAa,MAAM,cAAc,gBAAgB,MAAM;AACvD,WAAS,MAAM,mCAAmC,gBAChD,OACA,EAAE,OAAO,MAAM,CAChB;AACD,kBAAgB;KACd;AAEJ,QAAO;;;;;;;;;;;AA2BT,eAAsB,cACpB,OACA,SACA,MACyB;AACzB,KAAI,QAAQ,WAAW,EAAG,QAAO,EAAE;CAEnC,MAAM,gBAAgB,MAAM,iBAAiB;CAC7C,MAAM,OAAO,MAAM,QAAQ,QAAQ;CAGnC,MAAM,aAAa,QAAQ,MAAM,GAAG,cAAc;AAElD,OAAM,cAAc;CAGpB,MAAM,UAAU,IAAI,MAAM,WAAW,OAAO,CAAC,KAAK,MAAM;CACxD,MAAM,YAAY,WAAW,KAAK,MAAM,EAAE,QAAQ;CAElD,MAAM,SAAS,WAAY,SAAS;EAClC,WAAW;EACX,SAAS;EACT,YAAY;EACb,CAAC;CAQF,MAAM,UALS,MAAM,OAAQ,OAAO,EACd,OAIY,QAAQ;CAG1C,MAAM,SAAS,WAAW,KAAK,QAAQ,OAAO;EAC5C,GAAG;EACH,OAAO,OAAO,GAAG;EAClB,EAAE;AAGH,QAAO,MAAM,GAAG,MAAM,EAAE,QAAQ,EAAE,MAAM;AAExC,QAAO,OAAO,MAAM,GAAG,KAAK"}