Knowledge Base, Now A-Grade
A bot that answers from your documentation is only as trustworthy as the passages it pulls. We rebuilt the Knowledge Base around a single, authoritative retrieval engine — so the sources your bot grounds its answers in are the right ones, every time.
The problem this solves
Retrieval is the invisible half of a knowledge bot. When it quietly returns the wrong passages, the bot sounds confident and is wrong — the worst possible failure. Worse, the preview you tested with didn't always match what ran in production, so a bot that looked perfect in the Playground could behave differently for real customers. That gap made knowledge bots hard to trust.
How it works
- One authoritative engine. The Playground, readiness checks, and your live bot all query the same retrieval engine. What you preview is what your customers get — no surprises.
- Per-connection control. Tune each knowledge source independently: choose its embedding model, how many passages to retrieve (top-K), and the similarity threshold for what counts as relevant.
- Right sources, every time. Sources are isolated so different embedding models never collide, which means retrieval consistently returns the passages that actually match the question.
- Better grounding. Retrieved context is formatted so the model reads and uses it well, instead of ignoring it.
What this means for you
Your knowledge bots become dependable. You can test in the Playground and trust that production behaves identically, tune retrieval per source to balance precision and recall, and ship answers that are genuinely grounded in your content. Fewer hallucinations, more citations you can stand behind.
Getting started
- Open Knowledge Base.
- Pick a connection and set its embedding model, top-K, and threshold.
- Test in the Playground — and trust that production matches.
Cleaner retrieval, more accurate answers.