How we reduce wrong answers
We built BizPilot so your chatbot stays grounded in your business — and says 'I don't have that' instead of guessing.
Multi-layer confidence
We don't rely on a single number to decide if an answer is safe. Every knowledge-based reply is scored using several signals: how well the retrieved text matches the question (similarity), whether it actually answers the question (semantic check), whether multiple pieces of content agree (chunk agreement), and how relevant it is to your industry (vertical relevance). Only when the combined confidence is high do we show that answer; otherwise we fall back to capturing the lead or asking for clarification instead of guessing.
Hybrid search
We combine vector (semantic) search with keyword-style search so we can match both meaning and exact terms (e.g. service names, hours, pricing). That means we pull the right snippets from your site and docs more reliably than with vectors alone, which reduces wrong or off-topic answers.
Answer validation
Answers are checked against the sources we retrieved. We avoid making up details that aren't in your content — so the bot is designed to stay within what it knows from your knowledge base and avoid making up details.
What you see in the dashboard
In conversation detail we show confidence and source information when available, so you can see how the system is behaving and that responses are constrained to your content.
Our commitment
We use these checks so that only a small share of knowledge answers need to trigger our low-confidence fallback (e.g. “I don't have that information” or handing off to you). We continuously tune thresholds and add safeguards to keep that share low while still answering as many valid questions as possible.