Back to the room

Pass 3 of 3 · Deep dive of PHAROSPass 1 · OverviewPass 2 · How it works

Code Observatory / Source

Dense search with pgvector

Cosine distance through the HNSW index, filtered to one embedding model, with an iterative scan on pgvector 0.8 or later.

PINNED SOURCE Copied from the repository at this commit. Source, not an execution.

PHAROS at 975a38dMIT license

Retriever._densebackend/app/codeintel/retrieval.py, lines 370–405Open source
    def _dense(self, qvec: list[float], filters: Filters, limit: int) -> list[tuple[str, float]]:        assert self.embedder is not None        model = self.embedder.model_id        if not self.is_pg:            rows = self.db.execute(                select(DocumentChunk.id, DocumentChunk.embedding).where(                    DocumentChunk.embedding_model == model, *self._filter_clauses(filters)                )            ).all()            scored = [(str(cid), _cosine(qvec, emb)) for cid, emb in rows if emb]            scored.sort(key=lambda x: (-x[1], x[0]))            return scored[:limit]        # Higher ef_search keeps recall up when filters discard candidates        # after the HNSW scan (pgvector post-filters). pgvector >= 0.8 can        # also keep scanning until enough rows pass the filter, which matters        # once one database holds many repositories; results are re-sorted        # below because the relaxed order is only approximately sorted.        self.db.execute(text("SELECT set_config('hnsw.ef_search', :ef, true)"), {"ef": str(max(100, limit * 4))})        if _supports_iterative_scan(self.db):            self.db.execute(text("SELECT set_config('hnsw.iterative_scan', 'relaxed_order', true)"))        distance = DocumentChunk.embedding.op("<=>", return_type=Float)(            bindparam("qvec", qvec, type_=EmbeddingVector())        )        stmt = (            select(DocumentChunk.id, distance.label("distance"))            .where(                DocumentChunk.embedding.isnot(None),                DocumentChunk.embedding_model == model,                *self._filter_clauses(filters),            )            .order_by(distance)            .limit(limit)        )        rows = sorted(((str(cid), 1.0 - float(dist)) for cid, dist in self.db.execute(stmt)),                      key=lambda x: (-x[1], x[0]))        return rows

Where this comes from

Commit
975a38d5d17839c26ee219295a5be6ac66b6d910
File
backend/app/codeintel/retrieval.py
Lines
370–405
Git blob
884729a39ae0b3d340ad83f75946677dcddf3791
License
MIT, © 2024–2026 Pharos Contributors. Notice

Read it on GitHub at this commit

Check it yourself

From a clone of the repository, the first command prints the blob above and the second prints these lines.

git rev-parse 975a38d:backend/app/codeintel/retrieval.py
git show 975a38d:backend/app/codeintel/retrieval.py | sed -n '370,405p'