AI at Workfrom Acuity AI Advisory

Case studies · Architecture practice, Ireland

Building a private document knowledge base with Claude

By Ger Perdisatt · Reviewed

An architecture practice in Ireland had decades of project knowledge in planning applications, project sheets and past bids, and wanted any AI used on those documents kept on its own premises. We used Claude to build a private knowledge base and the software that runs it: documents dropped into a shared folder are split into passages and indexed, and staff ask questions in plain English and get an answer that names the source document, or "Not recorded" when the evidence is not there. It runs on a small computer in the firm's office using an open AI model, not Claude; Claude was used to build it. It is a proof of concept: more than 900 of the firm's documents were indexed in a single working session.

Proof of concept Running on the firm's own hardware. Not yet in day-to-day use.

Client identity withheld to respect confidentiality.

The starting point

When a bid or a client needed evidence of past work, someone had to remember which project it was and go looking through file shares for the planning application, project sheet or past bid that held it.

The practice also wanted any AI used on its documents to stay entirely on its own premises, with no cloud service involved.

What we implemented

Claude wrote the software: the part that watches a shared folder, reads each new document, splits it into passages and indexes them, and the part that answers questions from those passages.

The answers come from an open AI model running on a small computer in the firm's own office. Claude does not see the firm's documents or answer the questions. We chose this set-up because the firm wanted nothing to leave the building.

Every answer has to quote the passages it relies on and name the document. If nothing in the indexed documents supports an answer, it says "Not recorded" rather than guessing.

Using it, step by step

  1. 1What the person supplies

    A folder of project documents: planning applications, project sheets, reports and past bids. Staff add new documents by dropping them into the folder.

  2. 2What happens next

    Each document is read, split into passages and indexed automatically. A person asks a question in plain English and gets an answer with the passages and document names behind it.

  3. 3What they check before using it

    Before relying on an answer, for example in a bid, the person opens the named document and checks the passage. "Not recorded" means the indexed documents do not hold the answer, not that the firm has never done the work.

Example output

A question, its answer with document references, and the passages behind it. A follow-up question the documents cannot answer gets "Not recorded".

A fictional question and answer screen. The question asks whether the practice has designed a school with a timber frame. The answer names a fictional school extension from 2019 and cites a project sheet and page 3 of a planning report. Below, the two supporting passages are quoted with the matching words highlighted. A follow-up question about the roof U-value is answered Not recorded.
Illustrative example using fictional data. A plain-English question answered from the firm's own documents, with the supporting passages and a "Not recorded" answer.

Human checks

  • Answers name their source documents, so staff can check them before use.
  • When the documents do not support an answer, the system says "Not recorded" instead of guessing.
  • Only documents in the shared folder are indexed. The firm decides what goes in.
  • Nothing leaves the building: the AI model runs on the firm's own hardware.

Results

Proof of concept Running on the firm's own hardware. Not yet in day-to-day use.

  • Measured

    900+ documents loaded and indexed live in a single working session at the practice.

  • Estimate

    Seconds, not hours to find project evidence that used to be dug out by hand. Not timed against a manual search.

  • Observed

    Within minutes the team was asking questions of its own material during the session.

Measured: counted or timed. Estimate: worked out from the old routine, not timed. Target: what the client is aiming for. Observed: seen in use, not a number.

Limitations

  • It is a proof of concept and is not yet in day-to-day use.
  • An open model running on a small office computer is generally less capable than the largest cloud models such as Claude. It is good at finding and quoting passages, and weaker at reasoning across many documents at once. That is the trade the firm chose for keeping everything in-house.
  • Answers are only as complete as the documents in the folder.
  • The time saving is an estimate, not a measured result.

Which Claude product did what

Software built with Claude
Claude wrote the software that indexes the documents and answers questions. An ordinary program that Claude wrote and tested. Once it is built it runs on its own, so no AI is involved when staff use it.
Not Claude: a local open model
Answers staff questions on the firm's own computer. Claude is not involved when staff use it. An openly available AI model running on the firm's own computer. It is not Claude and sends nothing to Anthropic or any other cloud service.
Claude Team
Not used in this work. Anthropic's business plan for Claude. Staff use Claude in a browser or the desktop app, share projects, and connect it to systems such as Microsoft 365. The workspace and its data belong to the business.

How Claude Team, Claude Code and software built with Claude differ

Other case studies

Interested in a private knowledge base for your business?

Tell us how the work is done now and who does it. We'll suggest a time for a short call. See also ways to start.

Talk to us about a private knowledge basehello@acuityai.co