Flow Lab

An assistant that answers from your policies or says it can't

The assistant answers staff from company policies and shows the exact paragraph, version and page under every answer. When a newer version replaces a rule, the assistant flags it, so staff see the current rule at once. If the answer is not in the documents, the assistant says so directly and lists the 3 closest sections.

I built this demo on my own, with no client behind it, so the freight company in it is fictional. The chat on the demo page shows the document search only. The language model's answers in the demo were recorded in advance on the same questions. Every figure on this page comes from the demo itself, so you can check each one.

Screenshot: An assistant that answers from your policies or says it can't
8PDF policies the assistant answers from
29 of 31right paragraph and fact found
9 of 9unanswerable questions refused
1 business dayproject review at no charge with your request
Project example · demo

Project goal. The demo answers common questions from a freight company's staff about work rules. For example, staff ask about hotel limits, mileage pay and what to do after a traffic accident. The answers are in 8 company policies in PDF. Some rules in these policies have already been updated with newer versions that now apply.

How the assistant works. The assistant answers only from the company's own documents. Each answer names the document, version and section the assistant relied on. The exact paragraph appears under the answer, so staff can check it against the source at once. A link in the answer opens the PDF on the right page, so staff can read the rule in the document itself. If a rule has changed, the assistant flags it and shows the old version with a note that it no longer applies.

Test results. After tuning, a separate AI assistant wrote 40 new test questions. It saw only the documents, not the assistant's code. For each question, the model answered from the 5 most relevant paragraphs in the policies. Of the 40 questions, 31 had an answer in the documents. For 29 of those 31 questions, the model found the right paragraph and the exact fact. The model correctly spotted all 9 questions the documents do not cover and declined to answer them. Every quote in the answers matched the policy text word for word. Each quote came from the current version of a policy.

For your project, I will set the assistant up on your policies and run the same test on new questions. The assistant is built in Claude Code, Anthropic's AI coding tool. AI writes the first draft of the code. I check the code and run every test myself. The first step is an assistant for part of your documents, so you can see how it answers from your own policies. You get a price once you describe the project. The timeline comes in your plan within one business day. You pay in stages, each one after you approve it. Work starts 2 business days after you confirm the quote. A quote holds for 7 days. I quote support after launch once you describe the project. Support is billed monthly, so you can cancel it month to month.

Your order includes the first month of support at no charge, with one check and up to 30 minutes of edits. This applies if you confirm the quote within 7 days.

Next step

An assistant for your policies.

Send a list of your policies, or 2 or 3 sample documents, and add one line on the questions staff ask most. Within one business day of your request, you get a free plan with a price and timeline. Any bug found within 30 days of delivery is fixed at no charge.