Knowledge and document chat

By default the AI answers from what it learned during training. Knowledge lets it answer from your documents instead — a policy handbook, a product spec, a stack of reports — and quote the parts it used.

Under the hood this is RAG (retrieval-augmented generation): when you ask a question, NUFI finds the most relevant passages in your documents and hands them to the model along with your question, so the answer is grounded in your content rather than guessed.

Guide: For a full walkthrough, see Build a knowledge assistant.

Two ways to chat with a document

For a one-off question, the quick path is enough — attach the file, turn on File Search, and ask. Read on for the agent path, which keeps your documents available across new chats and shares them with a team.

Build a knowledge agent

An agent is a reusable assistant (see Agents and skills). Turning on File Search and adding documents gives it a permanent memory of those files.

Once the agent exists and File Search is on, the knowledge section shows each document and an Upload for File Search button to add more:

Create first, then upload. For a brand-new agent the upload button is disabled with the note "Agent must be created before uploading files for File Search." Fill in a name and model, click Create, and the button turns on.

Ask questions

Select the agent and ask in plain language. Because retrieval happens per question, you can keep the conversation going:

Summarise this handbook's leave policy in three bullets.

How many days of paid annual leave do full-time employees get?

Does the remote-work section mention meeting-free days?

The AI pulls the relevant passages and answers from them — for the question above it would answer 18 days because that's what the document says, not a generic guess. When an answer is built from your files, NUFI shows which document it drew on so you can check the source.

Share knowledge with a team

Upload once, let your whole team ask. On a team's Knowledge tab you can share a document with everyone (or with a specific group), and it becomes shared document chat for the team.

See Teams and sharing.

What works well

• File types — PDF, plain text, Markdown, and common document formats. Text-based documents retrieve best; a scanned image of text won't unless it has a text layer.
• Structure helps — clear headings and short sections retrieve more accurately than one long wall of text.
• Split very large documents — several focused files usually beat one enormous file.
• Name files clearly — names show up next to the answer, so 2024-hr-handbook.pdf is friendlier than doc1.pdf.

Your admin sets the upload size and count limits — see Files and images for the typical defaults.

Tips for better answers

• Ask specific questions. "What's the refund window in section 4?" beats "tell me about refunds."
• Point at the document in your wording — "according to the handbook…" nudges the model to lean on the file.
• Add instructions to the agent like "Only answer from the uploaded documents. If it isn't there, say so." to keep it from filling gaps with guesses.
• Keep knowledge tidy. Remove outdated files so old text can't be retrieved by mistake.

Troubleshooting

• The answer ignores my document — make sure File Search is enabled on the agent and the file finished processing (its card is showing), and that you're chatting with that agent.
• It answers from general knowledge instead — the passage may not have been retrieved. Ask more specifically, or add an instruction to answer only from the uploaded files.
• "File too large" — over the per-file cap. Split or compress it.
• A scanned PDF returns nothing — it has no selectable text. Use a text-based PDF or run OCR first.
• A teammate can't use it — knowledge shared through an agent or the team Knowledge tab follows team sharing; check the audience on Teams and sharing.