What does weakEvidence mean in a recall response?
Understand when Unison returns weakEvidence:true and what to do about it.
My recall response has weakEvidence: true. What does that mean?
It means the brain searched but found nothing that scored above its confidence threshold (approximately 0.5). In plain terms: the brain doesn't have confident knowledge about what you asked. The contextMd field may still contain something, but treat it as low-confidence filler rather than authoritative memory.
Example response
{
"contextMd": "",
"weakEvidence": true,
"sources": []
}What you should do
Don't pad the prompt with weak evidence. If weakEvidence is true, the safest move is to omit the contextMd from your prompt entirely and let the model answer from its base knowledge. Injecting low-confidence context can actively make answers worse.
A good pattern:
const { contextMd, weakEvidence } = await recall(query);
const systemBlock = weakEvidence ? '' : contextMd;Why is it weak?
Three common reasons:
- The brain doesn't have relevant data yet. Ingest more facts or conversations related to this topic.
- The query is too vague. More specific queries surface higher-scoring matches.
- You're using
standardmode and the signal is spread across many small facts. Trymode=deep— it aggregates across more sources and may cross the threshold.
See Improving recall for practical steps.
How do I rotate my API key?
Cycle a compromised or stale Unison API key by minting a fresh one and revoking the old — with zero downtime and no data loss.
What's the difference between auto, standard, and deep recall modes?
Choose the right recall mode for your use case: auto for most situations, standard for speed, deep for thorough aggregation.