01 |HYBRID RETRIEVAL
The cheapest path that can answer. Escalating only when it must.
A repeated question whose facts have not changed reuses the earlier answer. Otherwise hybrid retrieval ranks candidates, and a fallback widens the search when that is not enough.
Natural language query
Embed the query
gemini-embedding-2
Lexical search
passages, entity names, facts
BM25
Vector search
passages, turns, audio, images
HNSW · COSINE
Graph expansion
entity → relates_to → entity
1 TO 2 HOPS FROM SEEDS
memories + facts
02 |WHAT YOU GET
Seven signals, one ranker. Every score is recorded on the trace.
Vector
Semantic similarity over embeddings stored on the same records.
Read more
Lexical
Keyword matching, so exact terms survive paraphrase.
Read more
Graph
Connected entities pulled in through typed relationships.
Read more
Section links
Section embeddings and links between documents surface related sections that read apart.
Read more
PageRank
Importance weighted by how entities connect.
Read more
Geo
A geospatial signal for facts that are tied to a place.
Read more
Trace features
The history of what proved useful in earlier answers feeds the ranking.
Read more
Scores on the trace
Per-signal scores are recorded on the trace, so a ranking can be inspected afterwards.
Read more
Sources on every answer
Recall returns the records behind each result, ready to cite.
Read more
03 |THE ROUTER
Four tiers, cheapest first. Each query takes the cheapest path that can answer.
Tier 1
Structured seed reads
Reads over typed facts seed the ranking. They feed the ranker and are never the final answer on their own.
Tier 2
Response cache, invalidated by entity
A repeated chat question whose facts have not changed reuses the earlier answer. A cached answer is never reused once a fact it cited is superseded.
Tier 3
Hybrid retrieval
The seven signals are fused by one ranker: semantic similarity with keyword matching, graph relationships and the history of what proved useful.
Tier 4
Escalated fallback
When hybrid retrieval is not enough, the fallback uses HyDE and a doubled candidate pool.
GET STARTED
Give your agents recall they can trust. Four tiers, seven signals and a trace for every ranking, on one SurrealDB engine.
