Tallowby Freight · Policy assistant

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

Staff ask in their own words. The assistant searches the company’s PDF policies, quotes the exact passage, links to the page, skips versions that were replaced, and says so plainly when the documents don’t cover a question.

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. Code, version, effective date and status come from the document’s own control box.

Split by section

Numbered headings (4., 4.2) become passages. Each one keeps its page number, so a link opens the right page.

Search in staff’s words

A hand-made list of everyday words (petrol → mileage, nicked → stolen), and about a thousand sample questions a model wrote for the passages at import. Typos are matched to the nearest known word.

Quote, don’t guess

The answer is the passage itself. If the best match is weak, or shares too few of the question’s words, the assistant says it can’t find it.

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Superseded versions are used only when a question is about the past.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, both measured below. The chat on this page is the search step alone, running in your browser: no server and no keys, so it opens offline. It is careful rather than clever. 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. In a client project the next step is a language model that reads the five passages the search found and answers with a quote and its section. The test includes a recorded run of that step.

Checked, not promised

40 questions written separately, in another assistant session that saw the documents but not the code, and checked to be unlike the questions used for tuning. The search was frozen before they were first run. They include traps: topics no document covers, and rules that changed between versions. The left column is the search in your browser. The right column is a recorded run of the full pipeline: for each question the same search picked five passages, and a language model answered from those five only. Every quote it gave is checked here, word for word. Every miss on both sides is listed below.

What a real 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

SharePoint, Google Drive, Notion or a shared folder. A new version replaces the old one on upload, with no re-typing.

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

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

Bigger libraries

Past a few hundred pages, search adds embeddings next to keyword ranking and re-ranks the best passages. The test set works the same way, so the gain is measured, not assumed.

Scans and tables

Scanned PDFs go through text recognition first. Tables are kept row by row, the way the London hotel limit is kept here.

Monthly upkeep

A log of what people asked and what the documents didn’t answer. The test set grows each month, and gaps go back to the document owner.