Skip to content
PI Square
Book

Knowledge assistants

Answers from your own documents, with the source attached, for your staff and your customers.

drive · intranetwikichunkembed?asked byassociatekeywordmeaningindex · who may see each chunkrerankmodelq: refund, no receipt?yes, with proof ofpurchase, as credit[1] returns procedure 4.2[2] store policy p.3the answer, with its sources
  1. 1Connect sources
  2. 2Chunk and embed
  3. 3Search: keyword and meaning
  4. 4Rerank
  5. 5Answer with sources
Try it

A knowledge assistant answers questions from your company's own documents: policies, manuals, contracts and product data. It uses retrieval-augmented generation (RAG): it finds the passages that answer the question, writes the answer from them, and shows where each part came from. PI Square builds them so people only get answers from documents they are already allowed to open.

  1. Your documents

    SharePoint, Google Drive, wikis, PDFs, databases

  2. Your cloud

    Indexed with permissions

    who may see each passage

  3. The right passages found

    by meaning and by keyword

  4. Answered from them

    with every source linked

  5. Staff and customers ask

    in Teams, Google Chat, your intranet or your app

What it's for

  • Retail

    Store teams ask how to handle a return or a stock count, and get the answer from the current procedure.

  • Financial services

    Staff find the right clause in credit policy or product terms, with the page it came from.

  • Professional services

    Proposal teams reuse past answers to tenders without searching shared drives.

  • Customer service

    Agents and customers get the same answer, from the same source.

How we keep it safe

  • Access rights are carried into the index, so an answer never draws on a document the person can't open.
  • Every answer links to its sources. When nothing relevant is found, the assistant says so.
  • Personal information can be masked before anything reaches the model.
  • Before launch, answers are tested against real questions from your team, and tested again after every change.

Built with

  • Vertex AI Vector Search
  • Gemini
  • Claude
  • OpenAI
  • Python
  • FastAPI
  • Cloud Run

Where we've built this

  • A regulated consumer AI product: an AI companion for mental health and wellness, with every model call checked before it runs.
  • A southern African retail group: an assistant that asks employees follow-up questions in Google Chat and builds a complete brief from their answers.

How an engagement runs

  1. Understand

    We pick the questions people ask most and the documents that answer them, and test the assistant on a sample.

  2. Prove

    One assistant for one team, on the real documents, with the sources of every answer on show.

  3. Embed

    Your team owns the sources, permissions and evaluation questions, with an agreed way to maintain the assistant. Additional teams and sources are a separate decision.

The full approach

Questions about knowledge assistants

What is retrieval-augmented generation (RAG)?
RAG makes a language model answer from your documents. The system first retrieves the passages that match the question, then the model writes the answer from those passages and cites them. Your documents stay in your cloud, and the model is not retrained on them.
Will the assistant make things up?
Language models can, so we design against it. The assistant answers only from the passages it found, shows its sources, and says when it can't find an answer. Before launch we test it on real questions from your team and check that each answer matches its source.
Can it respect who is allowed to see what?
Yes. Permissions from SharePoint, Google Drive or your own systems are stored with each passage in the index, and every search is filtered by the person asking.
How is retrieval accuracy tested before deployment?
We compile an evaluation set of representative questions paired with verified passages from your documents. We measure retrieval recall and answer fidelity, checking that every answer is grounded in its source and that unanswerable questions are flagged.
Does company data remain private and secure?
The engagement defines where documents and indexes are hosted, who can access them and how model providers process data. Retention and training terms depend on the selected services and configuration; we review those requirements before using business data.

The other five

Tell us what you're working on

Keep this general; leave out sensitive or confidential information.

Or email nikhil@pisquare.ai