Why is the agent using an old answer?

Open Workspace → AI Agent → Learning → Preview. This runs the agent's real search and shows the score behind each result broken into its parts — which is the answer to why it keeps reaching for one answer over another.

This article is for tuning an assistant that is already working. If the agent is answering nothing at all, start with How do I set up the AI agent?.

Reading a preview

Type a question the way a customer would ask it and press Search. Each result shows:

• Similarity — how close the stored answer is to the question asked.
• Recency — how recent it is, on a curve rather than a cliff.
• Score — the two combined, which is what actually decides the order.

The formula is shown under the results:

Score = semantic weight × similarity + (1 − semantic weight) × recency weight

So an old answer wins when its similarity is high enough to outweigh its age. That is usually a sign the newer answer is worded further from how customers actually ask, not that the ranking is broken.

No matches means the agent would retrieve nothing for that question — a gap to fill with a document or an approved answer.

The four settings

Learning → Settings holds the levers:

• Half-life (days) — where an answer's recency weight halves. Default 90. Lower makes the agent forget faster; higher keeps old answers competitive.
• Semantic weight — the share of the score that comes from similarity, the rest being recency. Default 0.7. Raise it to favour the closest answer whatever its age; lower it to favour recent ones.
• Stale after (days) — past this an answer is never retrieved, whatever it scores. Default 540. This is the cliff behind the curve.
• Max results — how many answers the agent sees per turn. Default 3.

Change one at a time and re-run the preview. Two changes at once and you cannot tell which one did the work.

Half-life tuning

Learning → Tuning does the measurement for you. It re-asks each approved answer's own question and reports, for a range of half-life values:

• Precision@1 and Precision@3 — how often the right answer comes back first, or in the top three.
• Staleness — how much of what comes back is old.
• Coverage — how many of your sampled questions found anything at all.

Read where precision@1 stops improving against where staleness starts rising. They move in opposite directions, and the half-life worth picking is the one just before precision flattens.

Tuning only samples approved answers, so a workspace with a handful of them gives a noisy result. Wait until you have a real set before trusting the curve.

I changed the settings and nothing moved

Retrieval settings affect what the agent finds, not what it has. If a question has no good answer stored, no setting will produce one — add a document or approve an answer instead. The preview's No matches is the signal for that.