A self-initiated demo by Flow Lab. Tallowby Freight is a fictional company. How it was built and what it costs

Tallowby Freight · Policy assistant

Ask the handbook. Get the passage, the page and the version.

Staff ask in their own words. The assistant searches the company’s PDF policies and shows the passage it finds word for word, with a link to the page. It uses superseded versions only for questions about the past. If the answer isn’t in the documents, it says so plainly.

This is a technical demo. The chat runs the document search only. Answers from search plus a language model were recorded in advance and are shown under “Checked, not promised.”

Ask a question

Try one of these, or type your own. Casual wording is fine.

How it works

Five steps, and each one can be checked on its own. In a client project the same steps can be connected to SharePoint, Google Drive or a shared folder.

Read the PDFs

Text is taken page by page. Document code, version, effective date and status come from the document control block in each PDF.

Split by section

Each numbered section (4., 4.2) becomes a passage. The passage keeps its page number, so a link opens the right page.

Match everyday words

Search draws on a hand-built list of everyday words (gas → mileage) and about a thousand sample questions a model wrote for the passages when the documents were prepared. Typos are matched to the nearest known word.

Quote, don’t guess

The answer is the passage itself. If the passage found fits the question poorly, the assistant says it found no answer.

Check the version

Superseded versions are used only when a question is about the past. When a rule changed, the old wording is shown with the date it stopped applying.

Two ways to answer, tested below: search alone, and search followed by a language model. Once the page has loaded, the search runs in your browser, with no server calls and no API keys. It is tuned to be cautious. In the test below it says “I can’t find this” for many questions that do have an answer, and then lists where the answer might be. The second way, the one a client project uses, adds a language model: it reads the five passages the search found and answers with a quote and its section.

Checked, not promised

The test set has 40 questions the search never saw during tuning. An AI assistant wrote them in English in a separate chat, with access to the documents but not the code. For the Russian version, they were rewritten in everyday Russian in another separate chat, also without the code. The search was left unchanged until the first test run. The questions include traps: topics no document covers, and rules that changed between versions. The left column shows how the search in your browser answers. The right column shows answers recorded in advance from a language model that used only the five passages the search found. This page checks every quote the model gave against the document, word for word. Every miss on both sides is listed below.

What a client project adds

The demo shows the part that decides whether staff trust the answers. A client project adds the rest.

Your sources, kept in sync

The assistant connects to SharePoint, Google Drive, Notion or a shared folder. When you upload a new version of a document, it replaces the old one, and nothing has to be copied over by hand.

Written answers with citations

A language model writes two or three sentences from the quoted passages. If it can’t cite a passage, it doesn’t answer.

Where people ask

An intranet page, Slack, Teams or Telegram. Access by role, so drivers don’t see finance-only documents.

Bigger libraries

Past a few hundred pages, the search adds meaning-based matching (embeddings) to keyword ranking, then re-ranks the best passages. The same test set runs against it, so the gain is measured, not assumed.

Scans and tables

Scanned PDFs go through text recognition first. Tables are stored row by row. In this demo, the London hotel limit is one of those rows.

Monthly support

A log of what people asked and what the documents didn’t answer. The test set grows each month, and topics with no answer are passed on to the document owner.