Every correction
compounds.
When the bot says something wrong, an admin clicks "Coach" on the message, types the correction, and the guidance is embedded into a vector store. The next time a similar question comes in — minutes, days, or months later — that coaching is retrieved by semantic similarity and injected into the prompt. Quality compounds without retraining the model.
Coach once, apply forever.
Built on a ChromaDB vector store — every coaching entry is embedded at write time and retrieved by semantic similarity at chat time.
One-click coaching
Every assistant message in the admin conversation viewer has a "Coach" button. Click it; type what the bot should have said (or what rule it should have followed); save. The entry is stored with the original query, the original response, and the new guidance.
Semantic recall
At chat time, the orchestrator searches the response_guidance collection for entries semantically similar to the incoming customer query. The top three matches are injected into the system prompt as "COACHING FROM ADMIN REVIEWS." The bot reads them as instructions.
Calibrated against ground truth
Coaching entries that fire often (high hit count) and lead to better outcomes (better customer ratings, fewer escalations) are weighted up. Entries that misfire or get rolled back are surfaced for review.
Why this matters.
Most LLM platforms can only improve the bot by editing the system prompt or fine-tuning the model. Both are slow, hard to roll back, and require an engineer.
No engineering required
Anyone with admin access can teach the bot. The coaching entry is plain English, the recall is automatic, and the change is live within minutes (cache TTL of one minute).
Fully reversible
Bad coaching is a one-click rollback. The entry is archived, the embedding is removed, and the bot stops applying the guidance immediately. Compare this to fine-tuning, which requires retraining and re-deploying the model.
Coaching meets Agent Workshop.
The Coaching System and Agent Workshop work together to give you complete control over your AI's behavior.
Response guidance per persona
Coaching entries apply across all six agent personas. When Sales, Support, or After-Hours receives a similar query, the same guidance fires. Train once, improve everywhere.
Workshop + Coaching = full control
Agent Workshop sets the global personality, tone, and capability boundaries. Coaching fills in the edge cases the system prompt doesn't cover. Together they eliminate the "but what about when..." gap.
Simulation Studio testing
After adding coaching entries, test them in the Simulation Studio. Run scenarios against draft config to see how new guidance interacts with the persona prompts before going live.
Coaching System, running on your data.
30-minute walkthrough on your real catalog. No slideware, no demo data.