These controls are for platform administrators in AI Agent. Keep tuned values unless realistic Knowledge tests show a problem.
Model and credentials
- Generation uses OpenAI gpt-6-luna, with no model escalation. Gemini controls are retired.
- Existing Knowledge uses text-embedding-3-small, 1,536 dimensions. Changing embeddings requires Knowledge/catalog reindexing and renewed review of platform support publication.
- A chat-model change alone does not require embedding changes.
- The OpenAI key serves generation, embeddings and transcription. An environment key takes precedence over a saved key.
- Firecrawl handles website imports; files and manual sources do not require it.
Starting values
- Match threshold: 0.18. Filters passage similarity. Lower values return more passages; customer retrieval retries at 0.18 if a higher cutoff finds none.
- Confidence threshold: 0.65. Lower-confidence customer answers receive Luna review for clarification or handoff. This score is not a measured accuracy percentage.
- Match count: 6. More passages can increase latency and token use.
- Conversation lookback: 10 messages. Controls recent customer context, not merchant Assistant memory or queued tasks.
- Firecrawl concurrency: 2. Controls website ingestion, subject to provider limits; it does not batch live conversations.
Verify a change
Test realistic Arabic/English questions, follow-ups, misspellings, missing facts and handoffs. Inspect matched sources as well as wording. Do not change embeddings just to improve an individual answer.
Still need help?
Tell us what you were trying to do, the channel involved, and what you expected to happen.