Every answer follows the same discipline: find the relevant knowledge first, then answer from it. No knowledge, no invention.
The pipeline, in plain terms
- Understand the question. A fast pre-check filters out greetings and chatter (you are never charged for those) and rewrites the question for search.
- Search your knowledge. OnyxBot runs a hybrid search across everything you connected: meaning-based matching for paraphrased questions, keyword matching for exact terms like error codes.
- Keep only what's relevant. Weak matches are dropped rather than stuffed into the prompt. If nothing survives, the AI says so instead of guessing, and inside a ticket, coming up empty twice in a row triggers escalation to your staff.
- Answer and remember. The reply stays clean in Discord, but OnyxBot keeps a record of what the search found, what made it into the prompt, and which pages the model said it cited. Open the answer on Home, then Usage and its Sources section shows exactly what it read, next to the exact cost of the answer. Context is kept for follow-ups.
What makes answers better
- Connect the pages members actually ask about.
- Approve learned answers from resolved tickets; real resolutions are the highest-signal knowledge there is.
- Use custom instructions for tone and boundaries, and knowledge sources for facts.
- When an answer is wrong, read its sources before you edit anything. The fix is usually one page, and the sources tell you which.
Honesty over confidence
OnyxBot is deliberately tuned to prefer "I don't know, let me get you a human" over a fluent wrong answer. In a support product, a confident hallucination is worse than an escalation.