Project documents: AI answers about cover, straight from your policy wording

A price is only half of what a broker or customer asks about. The next question is nearly always about cover. Is flood in the basement included? What is the excess on a burst pipe? How long do I have to make a claim?

Until now, an AI client connected to Ptally could price a quote and explain how the price was built. It could not answer those questions, because the wording lived somewhere else. Project documents fix that. The policy wording now sits in the same project as the model that prices it.

What it does

You upload the documents behind a product to its project in Ptally. That might be the policy wording, the product guide, the IPID or your underwriting guidelines. PDF, Word, text and Markdown are accepted.

Ptally reads each document and splits it into passages at its headings and numbered clauses. Every passage is labelled with its section and page. AI clients connected over MCP can then search those passages and read around them.

Ask Claude whether flood damage to a converted basement is covered, and it searches the wording. It comes back with the exclusion, the contents exclusion that goes with it and the insured perils. Its answer cites each one by section and page, so anyone can check it.

Three tools, all read only

Three MCP tools work with project documents:

• list_docs lists every document on a project, with its status. A document is indexed, still indexing, or failed with a reason.
• search_docs finds the passages that answer a question. Pass the version that priced a quote, and the search uses the documents that apply to that version.
• get_doc reads the passages either side of a hit. It is for a clause that is cut off or refers to a defined term.

Uploading and deleting stay in the dashboard. An AI client can read your documents, but it cannot change them.

Answers you can check

The weak point of any AI answer about cover is confidence without a source. We have built the tools to push the other way.

The MCP server tells clients to answer only from the passages returned. Each passage they use must be cited by title, section and page. When nothing returned answers the question, the client is told to say so rather than guess. Each search result also carries a similarity score. A low score is a signal that the documents probably don't cover the question.

None of this replaces the wording itself. It points the reader to the right clause quickly, and shows them which clause the answer came from.

How search works

Each passage is scored two ways. One score measures how close its meaning is to the question, using a small embedding model. The other is a keyword match. The two are blended 70 to 30.

The keyword half matters more than it looks. Policy wordings are full of exact terms: clause numbers, defined terms, section names. Someone asking about "Section 4" or "unoccupied" wants that exact clause, not the one that sounds most similar.

Where the documents go

Extraction, indexing and search all run on the Ptally server that holds your projects. The embedding model is small and open source, and it runs on that server's CPU. No document text is sent to an outside service to be indexed.

When an AI client searches, it receives the passages that matched, and nothing else. For a self-hosted install, that means the whole index stays in your own cloud.

Try it before anyone else does

Each project page now has a Project documents card. It shows what is uploaded, whether each document is ready and how many passages it holds. Try it now lets you type a question the way a broker or customer would. You see the passages an AI client would be given to answer it.

That is the quickest way to judge a wording before you rely on it. If the right clause doesn't come back for an obvious question, the document may need clearer headings, or a scanned PDF may need OCR first.

Two directions, one project

A Ptally quote already drives the policy documents built from it. Endorsement steps choose the schedules and clauses, and a template fills them from the response. That is covered in building policy documents from the quote response.

Project documents go the other way. They bring the wording in, so AI clients can answer questions about it. Together, the price, the forms that go with it and the wording that explains it all live in one project.

Getting started

Open a project, drop the wording on the Project documents card and wait for it to show as ready. Then ask your AI client a question about cover. The full reference, with a real search over a sample home wording, is on our project documents page. The search endpoints are in the API reference.