Skip to main content

​
Why hybrid

Vector search alone misses exact identifiers (file name (e.g. pg_hba.conf), a ticket number, or a specific concept) because embeddings blur precise strings into approximate meaning. Keyword search alone misses the meaning, so memloom runs both (plus an entity arm) inside the database, in a single SQL function (memloom_fuse), and fuses the rankings. There’s no reranker service and no second round trip: one query, one fused result list. All arms rank saved memories and context chunks together in one pool, so “search my memories” and “search my files” are the same operation.

​
The three arms

The entity arm has an abstention gate: it only fires when the query is clearly about a known entity (query ↔ entity cosine ≥ 0.60). A vague query anchors to nothing and the arm contributes nothing, so it never drags in unrelated results. The keyword arm abstains the same way: no lexical match, no votes.

​
Reciprocal-rank fusion

Each arm produces a ranking; fusion combines ranks rather than scores (cosine distances and FTS ranks aren’t comparable):
Default weights: vector 1.0, keyword 2.0, entity 0.5. Keyword is up-weighted because it abstains when it has nothing, so its votes are cheap and precise; the entity arm is half-weighted so it works as a tiebreaker between arms rather than a winner on its own (both eval-tuned on a real corpus). A result that two arms agree on beats a result that tops one arm alone. Each recall result carries both numbers:
  • rrfScore: the fused rank score. This is the order results arrive in.
  • similarity: the raw cosine signal alone, useful as a “how semantically close” query is to the result.

​
What makes results good in practice

  • Exact identifiers are excellent queries. The keyword arm makes memloom recall "pg_hba.conf" precise in a way pure-vector systems can’t be.
  • Chunks embed their breadcrumb. A markdown chunk is stored as Guide > Setup > Postgres\n\n<text>: the heading path is prepended to the text, so both the vector and the keyword arm see the heading context along with the paragraph.
  • Index your store. The entity arm only works over what memloom index has processed. An unindexed store still recalls fine on the other two arms.
  • Stale never surfaces. Every arm filters status = 'active', so superseded versions of a belief can’t shadow the current one.
In offline mode (no API key) the vector arm runs on deterministic hash embeddings, which makes texts sharing words score as similar, which is a real but low signal compared to cloud embedding. The keyword arm works at full quality offline; the entity arm needs indexing, which needs the LLM. For production recall quality, just configure a key here.