When the biology buried the exact name
Ask our memory for a record by its own name: 93% rank-1 on plain semantic search, 43% after the biological rescore. Why a smarter score lost the easiest query, and the fix we chose.
1 Nov 2026 · 1 min readThe easiest retrieval test there is: take a record, query it by its own declared name. It should come back first.
The measurement
30 fixtures, each queried by its own frontmatter name:
| ranking | rank 1 | top 3 |
|---|---|---|
| plain semantic | 28/30 (93%) | 30/30 (100%) |
| default, with biological rescore | 13/30 (43%) | 22/30 (73%) |
The default path, the one every agent and the UI use, lost more than half of all name lookups.
One live example
Query: “newest suno august 2026”.
plain semantic the right file at ranks 1, 2 and 3 similarity 0.551
default a review note similarity 0.431 → final 0.604
an archived plan similarity 0.374 → final 0.587
the right file: gone from the top 3
Why
The rescore multiplies in decay, importance and a “soul” weight: biology meant to favour what is alive and important. For exploration that is useful. For an exact lookup it is noise: a record marked low-importance or superseded cannot beat a high-importance planning document, even at a much lower similarity.
A smarter score lost the easiest query, because it ran on every query.
The decision
B1 biology moves AFTER the cross-encoder rerank, as a tie-break or prior,
gated by intent:
exact lookup, operational how-to → biology off
exploratory → biology on
Approved on 31 October 2026. Routing on intent before retrieval is established practice in industrial RAG 1; the new part is using it to switch a prior off. It lands in the next retrieval phase. Not yet shipped, so the 43% is still the live number on the default path.
The general lesson
If your ranker has a prior (recency, popularity, importance), measure it on the query type it should not affect. Exact-name lookup is the cheapest negative control there is: every record in your store is already a fixture.
- J. Schubert, M. Krug, J. Baumeister. Knowledge Retrieval with LLMs using Context-Specific Intent and Slot Classification. LWDA, CEUR-WSoriginalcorpus: ~/hybrid-retrieval
The measurements in this article come from these records in the running system.
hive chat #356— UNRESOLVED 1: exact-name retrieval 93% to 43%~/memory/retrieval-revise-blueprint-map-2026-10-31.md— decision B1: biology after the rerank, gated by intent
Our embedder and both rerankers we tested are sold as multilingual. Measured with paired Slovak and English queries, Slovak failed 55% of the time against 20%, and a cross-encoder gave three different documents the identical score.
Our unified search compared three different quantities in one column. The exact answer ranked below a weaker one, purely because of scale. What we measured, and what we changed.
Retrieval finds the nearest item. Search composes an answer from parts, the way a chemist plans a molecule backwards from what is on the shelf. The map we took from the chemistry, and what it changes in an agent.