Reciprocal rank fusion for agent recall

How RRF merges keyword and vector search into one ranked list by rank, not score - and how Seamless runs both legs in a single SQLite file with no vector database.

Reciprocal rank fusion (RRF) is a method for merging several ranked result lists into one: each document scores 1 / (k + rank) in every list that returned it, the scores sum, and the constant k (60, per the original paper and most implementations) damps the difference between rank 1 and rank 5. The point of fusing by rank instead of by raw score is that the input lists never have to agree on what a score means - BM25 values and cosine similarities live on incomparable scales, and RRF never compares them, only their orderings.

That property is why RRF is the standard fusion step in hybrid search, and why it fits agent memory retrieval unusually well. Agents query in two distinct modes: exact identifiers (tasks_claim, SEAMLESS_DATA_DIR, an error string) where keyword search wins and embeddings blur, and paraphrases ("why does the first query fail on cold start?") where vector similarity wins and keywords miss. A fused rank means neither mode has to be predicted in advance - a result that both legs rank highly beats a result only one leg loves.

How Seamless runs it

Seamless's recall - the single search entry point its agents call - runs both legs inside one SQLite file:

• Keyword leg: SQLite FTS5 full-text search over names, descriptions, and bodies.
• Semantic leg: cosine similarity over float32 embedding BLOBs stored in the same database, compared by brute force - no vector database, no ANN index, no second process.
• Fusion: RRF with k=60, over a candidate pool a few multiples of the requested limit so the fused order has room to differ from either leg's.

The fused score is the base order, not the final word: a favorite multiplies its score by 1.15, and a memory's decayed demand adds a utility nudge capped at +10%. Both are bounded reorderings of what fusion returned - neither can pull in a result the legs did not.

Brute-force cosine is a deliberate trade: a personal knowledge store holds thousands of memories, not millions of documents, and at that scale a linear scan is faster than the operational cost of an approximate index. The same shape would be the wrong call at corpus scale - this is a design for one developer's fleet of agents, not a search engine.

The degradation rule matters as much as the fusion: with no embedding provider configured - or a configured one unreachable - recall runs keyword-only rather than failing. A memory system that errors when an API key lapses is a memory system agents learn to stop calling.

For where recall sits among the three delivery paths (briefing, prompt-matched injection, explicit search), see recall; for what the store itself looks like, see memory & notes.