Demo build, no client
Handbook assistant: cites its sources, admits what it doesn't know
Ask anything about the employee handbook of Brightmoor Labs, a made-up company. Every answer links to the passage it came from. If the handbook doesn't cover your question, the assistant says so instead of guessing. The eval panel below shows how often it gets things right.
How the answers are produced (straight talk)
- RealRetrieval: a BM25 keyword search over the handbook's 53 paragraphs, running in your browser. It has a small hand-written synonym list (for example "PTO" finds "vacation").
- RealAnswers are extractive. The assistant picks the best one or two sentences and copies them word for word, so it can't make up a policy. No language model writes anything, and no API key or server is involved.
- RealThe "I don't know" gate refuses when the best passage scores too low, or when it covers less than half of the question's key words. Open "How this was found" under any answer to see the scores.
- SampleBrightmoor Labs and its handbook are fictional. The eval questions were written alongside the handbook, so they aren't an independent benchmark.
- AI useAI tools helped write the code, the fictional handbook and the copy. A client build would usually add embeddings and an LLM (OpenAI, Claude or a local model) to write fluent answers, while keeping the same citations and refusal checks.
Ask the handbook
Brightmoor Labs handbook
Eval panel
Each in-scope question passes only if the first citation is the right section and the answer contains the expected fact. Each out-of-scope question passes only if the assistant says it doesn't know. The questions run live in your browser when you press the button, and nothing is pre-recorded.
| # | Question | Expected | Got | Result |
|---|