AI Deflection
How AnswerLoops automatically answers questions using your knowledge base.
How it works
- A user posts a question in a connected Discord channel or Slack channel
- AnswerLoops embeds the question as a vector using your configured embedding model
- It searches the knowledge base for articles with high cosine similarity
- If the best match exceeds the confidence threshold, it generates an answer and posts it in the thread
- If confidence is below the threshold, it triggers human escalation
Connecting a GitHub repository is optional. With repos connected, the AI can also search your source code while drafting answers. Without repos, it answers from the knowledge base and prior resolved tickets alone — and is instructed never to invent project-specific details it can't find there.
Confidence threshold
Default: 0.8 (range 0–1)
- Higher threshold → AI answers less, escalates more → higher accuracy, more human work
- Lower threshold → AI answers more, escalates less → more coverage, higher risk of wrong answers
Configure per-integration in Settings → Discord or Settings → Slack.
Embedding model
By default: text-embedding-3-small (OpenAI). Articles and queries are both embedded with the same model — they must match for similarity scores to be meaningful.
If you change the embedding model, existing KB articles have stale embeddings. Re-import all articles to regenerate vectors with the new model.
KB match threshold
Articles below a cosine similarity of 0.45 are filtered out entirely before the LLM sees them. This prevents hallucination from low-quality context.