refactor(memory): run agent recall, memory tools and the server on the memory contract (P1.5)
- Orchestrator, memory tools, context loader and pattern write-back use a MemoryStore and a workspace scope instead of MindDB layers; the knowledge graph tools, cognify, memory linker and feedback handler are gone (graph hidden in v1, ADR-0002). - The server opens one mem0 store in <dataDir>/memory/, embedded by the in-process model on first use, and extracts facts about the user in the background with the model the user chose (memoryExtraction setting). - Legacy .mind memories are copied into the store once, in the background; the legacy files are only read. - Weaver consolidation, harvest, wiki, memory import, memory lanes, vector backfill and index reconciliation are removed; agent runs, loops, sub-agent results, connector fetch and the failed-turn capture write through the contract. Workspace sessions no longer hold a mind handle. - FileIndexer and the install audit move to waggle.db; recall date helpers and the reranker are copied into core. - The CLI REPL opens the same memory store.
This commit is contained in:
137
packages/core/src/recall/inprocess-reranker.ts
Normal file
137
packages/core/src/recall/inprocess-reranker.ts
Normal file
@@ -0,0 +1,137 @@
|
||||
/**
|
||||
* In-process cross-encoder reranker using @huggingface/transformers (ONNX).
|
||||
*
|
||||
* Cross-encoders take (query, doc) pairs and output a relevance score by
|
||||
* jointly attending to both — much more discriminating than vector dot
|
||||
* products. Use after RRF to rerank the top-K candidates from hybrid
|
||||
* search. Standard pattern in production RAG systems.
|
||||
*
|
||||
* Default model: Xenova/ms-marco-MiniLM-L-6-v2 — the canonical
|
||||
* SentenceTransformers cross-encoder, ~22MB on disk, ~30-50ms per pair
|
||||
* on CPU. Trained on MS MARCO passage ranking, generalizes well to
|
||||
* mixed-domain technical text.
|
||||
*
|
||||
* Alternative: Xenova/bge-reranker-base (~280MB, slightly higher quality
|
||||
* on out-of-domain queries). Set via `model` config.
|
||||
*
|
||||
* `@huggingface/transformers` is an optional peer dep — if not installed,
|
||||
* createInProcessReranker throws and the caller falls back to no reranking.
|
||||
*/
|
||||
|
||||
import path from 'node:path';
|
||||
import os from 'node:os';
|
||||
import { createCoreLogger } from '../logger.js';
|
||||
import { withTransformersModelLoad } from '../embeddings/transformers-model-load.js';
|
||||
|
||||
const log = createCoreLogger('inprocess-reranker');
|
||||
|
||||
export interface Reranker {
|
||||
/**
|
||||
* Score a single (query, doc) pair. Higher = more relevant.
|
||||
* Score scale depends on the model — for ms-marco-MiniLM it's
|
||||
* roughly [-10, 10]; relative ordering is what matters.
|
||||
*/
|
||||
score(query: string, doc: string): Promise<number>;
|
||||
|
||||
/**
|
||||
* Score N pairs sharing one query. Same-shape result as score() but
|
||||
* amortises the model invocation when supported.
|
||||
*/
|
||||
scoreBatch(query: string, docs: string[]): Promise<number[]>;
|
||||
}
|
||||
|
||||
export interface InProcessRerankerConfig {
|
||||
model?: string;
|
||||
cacheDir?: string;
|
||||
}
|
||||
|
||||
/**
|
||||
* Build a reranker backed by @huggingface/transformers. Throws if the
|
||||
* package isn't installed — caller is expected to catch and fall back.
|
||||
*/
|
||||
export async function createInProcessReranker(
|
||||
config?: Partial<InProcessRerankerConfig>,
|
||||
): Promise<Reranker> {
|
||||
const model = config?.model ?? 'Xenova/ms-marco-MiniLM-L-6-v2';
|
||||
const cacheDir = config?.cacheDir ?? path.join(os.homedir(), '.hive-mind', 'models');
|
||||
|
||||
log.info(`Loading in-process reranker: ${model} (~22MB first download)`);
|
||||
|
||||
const { AutoTokenizer, AutoModelForSequenceClassification } = await import(
|
||||
'@huggingface/transformers'
|
||||
);
|
||||
const { tokenizer, seqModel } = await withTransformersModelLoad({
|
||||
cacheDir,
|
||||
model,
|
||||
load: async (canonicalCacheDir) => ({
|
||||
// Cross-encoders need direct tokenizer + model access — pipeline API
|
||||
// doesn't expose the (text, text_pair) input pattern cleanly across
|
||||
// all transformers.js versions. Calling the model directly with
|
||||
// tokenized pairs is the stable path.
|
||||
tokenizer: await AutoTokenizer.from_pretrained(model, { cache_dir: canonicalCacheDir }),
|
||||
seqModel: await AutoModelForSequenceClassification.from_pretrained(model, {
|
||||
dtype: 'fp32',
|
||||
cache_dir: canonicalCacheDir,
|
||||
}),
|
||||
}),
|
||||
onQuarantine: () => log.warn(`Quarantined corrupt reranker model cache: ${model}`),
|
||||
});
|
||||
|
||||
log.info(`In-process reranker ready: ${model}`);
|
||||
|
||||
/** Score a single pair: tokenize, forward, extract logit. */
|
||||
async function scorePair(query: string, doc: string): Promise<number> {
|
||||
const inputs = await tokenizer(query, {
|
||||
text_pair: doc,
|
||||
padding: true,
|
||||
truncation: true,
|
||||
return_tensor: true,
|
||||
});
|
||||
const out = await seqModel(inputs);
|
||||
// ms-marco-MiniLM outputs a single logit per pair (1-class regression).
|
||||
// Other cross-encoders may output 2 classes — take logit[0] - logit[1]
|
||||
// as a relevance score in that case.
|
||||
const logits = out.logits ?? out[0];
|
||||
const data = logits.data as Float32Array | number[];
|
||||
if (logits.dims && logits.dims[logits.dims.length - 1] === 2) {
|
||||
return Number(data[0]) - Number(data[1]);
|
||||
}
|
||||
return Number(data[0]);
|
||||
}
|
||||
|
||||
return {
|
||||
async score(query: string, doc: string): Promise<number> {
|
||||
return scorePair(query, doc);
|
||||
},
|
||||
|
||||
async scoreBatch(query: string, docs: string[]): Promise<number[]> {
|
||||
if (docs.length === 0) return [];
|
||||
// Tokenize all pairs together for batch inference. Padding aligns
|
||||
// sequences so the model can process them in one forward pass.
|
||||
const queries = docs.map(() => query);
|
||||
const inputs = await tokenizer(queries, {
|
||||
text_pair: docs,
|
||||
padding: true,
|
||||
truncation: true,
|
||||
return_tensor: true,
|
||||
});
|
||||
const out = await seqModel(inputs);
|
||||
const logits = out.logits ?? out[0];
|
||||
const data = logits.data as Float32Array | number[];
|
||||
const dims = logits.dims;
|
||||
const lastDim = dims[dims.length - 1];
|
||||
|
||||
const scores: number[] = [];
|
||||
if (lastDim === 2) {
|
||||
for (let i = 0; i < docs.length; i++) {
|
||||
scores.push(Number(data[i * 2]) - Number(data[i * 2 + 1]));
|
||||
}
|
||||
} else {
|
||||
for (let i = 0; i < docs.length; i++) {
|
||||
scores.push(Number(data[i]));
|
||||
}
|
||||
}
|
||||
return scores;
|
||||
},
|
||||
};
|
||||
}
|
||||
Reference in New Issue
Block a user