Ask this library: AI chat over your documents

"Ask this library" is a front-end chat widget that answers visitors’ questions grounded in your documents — answers come from what your library actually contains, not the model’s general knowledge. It’s a familiar chat interface: message bubbles, a typing indicator, clickable example questions, and inline links to the source documents behind each answer. When a question is only loosely related to your library, it says so plainly rather than guessing. You configure it under Documents → Settings → Ask AI, and it requires your own API keys for both embeddings and chat.

FileDeck AI — Ask this library is part of the AI plan (from $119/year), which also includes semantic search and AI summary/tag suggestions. See pricing. AI features require your own AI provider API key.

What do I need before going live?

All three of these are required:

1. AI semantic search configured with your own embedding API key — set up on the AI search tab. See AI semantic search.
2. A chat provider API key — Anthropic (Claude) or OpenAI. You can optionally choose a chat model; the default is a fast, low-cost model.
3. The widget enabled so it appears on the front end.

How do I set it up?

1. Configure AI semantic search first and build the index at Documents → Settings → AI Index.
2. Go to Documents → Settings → Ask AI.
3. Choose your chat provider — Anthropic (Claude) or OpenAI — and enter its API key.
4. Optionally pick a chat model, or keep the fast low-cost default.
5. Enable the widget and save.

Can I set the starter questions?

Yes. The shortcode and block accept a prompts attribute — a pipe-separated list of example questions shown as clickable chips:

[filedeck_ask prompts="How do I request a refund?|What's covered by the warranty?|Where do I submit a claim?"]

Good prompts are the questions your visitors actually ask. They double as a hint about what the library can answer.

Can I see what visitors ask?

Yes — on the AI plan, with activity logging switched on. Documents → Activity Log → AI Questions shows the questions people put to the Ask box, split into answered and not found. The not-found list is the useful part: it’s a ranked view of what your visitors wanted that your library couldn’t answer, and each one has a one-click Create a document button that starts a new draft titled from the question. It turns Ask from a chat box into a content-strategy tool — every gap becomes the next thing you write.

The stored log of questions lives on your own server only — FileDeck never receives it — carries no IP address, and is automatically pruned after 90 days. Nothing is stored unless activity logging is on. (Answering a question does send its text and matching document snippets to the AI provider you configured, under your own account — see the privacy section above.)

What else does the chat key power?

The same chat provider key also powers AI summary and tag suggestions in the document editor, so once Ask AI is configured, you get writing help when adding or editing documents at no extra setup.

How is visitor privacy handled?

When someone asks a question, relevant text from your documents is sent to the chat provider to generate the answer. FileDeck offers an optional privacy setting that masks obvious personal data — emails, phone numbers, and card or SSN-like strings — before text is sent.

This masking is best-effort redaction, not a guarantee. If your library contains regulated data, use your own provider account with a data-processing agreement in place.

FAQ

Which chat providers are supported?

Anthropic (Claude) and OpenAI. In both cases you use your own API key, and you can optionally set the model.

Do I need separate keys for embeddings and chat?

Yes — the embedding key is set on the AI search tab and the chat provider key on the Ask AI tab. The widget will not go live until both are configured and the widget is enabled.

Are the answers really limited to my documents?

The widget is grounded in your library: it retrieves relevant document text via semantic search and answers from that, rather than from the model’s general knowledge.

Is the personal-data masking enough for regulated content?

No. It is best-effort, not a guarantee — for regulated data, use a provider account covered by a data-processing agreement.