agentsknowledgeevaluationsconversations
Faster replies, sharper retrieval, and evaluation fixes
Agents start answering sooner, low-relevance knowledge passages stop reaching answers, judged test cases get a fairer latency cap, and run history shows one dot per run.
Improved
- Faster replies — agents begin answering sooner across every channel, and a plain greeting no longer waits on routing work that cannot change the answer. Nothing to change on your side.
- Progress steps in dashboard chat and the web widget are now timed from when each step ran, so a step's duration reflects the work rather than when your browser received the update.
- Agent creation writes tighter core prompt files — the same coverage with less restatement, which leaves more room for knowledge and conversation history. See core prompt files.
- Fixtures authored with the in-dashboard assistant now flag a response field name that does not match the tool's real payload, so a mismatch surfaces before the case runs. See fixtures.
Fixed
- Answers drew on passages that matched only the common words in a question. Low-relevance passages are now filtered out before the agent sees them, so keyword and hybrid sources stay on topic. See retrieval modes.
- Test cases scored by the AI judge could fail the latency check on timing alone. Judged cases now get a longer default cap when you have not set one — set Maximum latency yourself to override it. See scoring.
- A case run with repetitions showed one dot per repetition in the run history strip. It now shows one dot per run.
- An agent that ended its reply with a follow-up question sometimes sent the whole reply twice.
- A remembered fact could appear twice in an agent's context when it had been stored both as a sentence and as a bare value.