Recall & Grounded Answers

'A confident answer with no source is worse than useless'

That line is the design thesis behind brainAskService, the surface that lets you interrogate the workspace Brain directly — "what did we decide about the Salesforce project?" — and get a written, cited answer, or an honest admission that the Brain doesn't know. Most memory products return raw chunks and call it retrieval; this one is required to either ground its answer in specific atoms or say so.

POST /v1/brain/ask
{ "question": "What policy controls outbound followups?", "scope": "workspace" }

Grounding on relevance, not just confidence

A retrieved atom must clear a relevance floor to be allowed to answer at all — a confident-but-topically-unrelated memory must never answer an unrelated question, which is exactly the bug that causes most systems to hallucinate-by-retrieval. Concretely: a hit needs real lexical overlap with the question (at least 0.12) and either a retrieval score above 0.34 or stronger overlap (0.2+) as a substitute. An atom that's confident but irrelevant is rejected outright, so an out-of-domain question correctly falls through to abstention instead of getting an answer that happens to sound plausible.

Diversified, not five near-duplicates

Retrieval blends TF-IDF over text fields with cosine similarity over embeddings, then diversifies the result set with MMR (maximal marginal relevance) at λ = 0.72 — strongly weighted toward relevance, but enough to stop a single dispatch context from being eaten by five near-identical atoms saying almost the same thing. An optional model-based reranker exists for a second, more expensive relevance pass, but it's off by default: the cheap MMR-plus-feature ranking already orders results well, so a per-query model call isn't spent unless an operator explicitly opts in.

Cited, or honestly silent

When enough evidence clears the floor, the answer is synthesized with inline [mem:id] citations tying every claim back to a specific atom. When nothing clears it, the service abstains rather than guessing — and even without a synthesis model configured, it degrades gracefully to a deterministic, fully-cited list of the grounding atoms themselves. Either way, you never get an answer you can't trace back to something the Brain actually knows.

Continue