Most organisations already have somewhere documents live: a shared drive, a document management system, an ERP with attachments. The useful integration is not another chat window — it is a search box inside the system people already open every morning.
The shape of the integration
- Create a folder on the appliance for the document set you want to expose.
- Issue an API key scoped to that folder. Keys are stored hashed and shown once, so copy it at creation.
- Send documents to the upload endpoint from your existing system, or let staff upload through the web interface.
- Call the query endpoint with a question and read back the answer and its citations.
- Render the citations as links into your own document viewer so users land on the exact page.
Key handling
Scope one key per integration, never per person, and revoke rather than reuse. A folder-scoped key cannot read folders it was not issued for.
Things worth knowing before you build
- With a local inference endpoint and no external integrations enabled, AI requests can remain on your network. A remote model provider or custom endpoint may receive prompts; verify the endpoint and egress policy.
- The current upload handler extracts and indexes a document synchronously before the request completes. Large uploads may take longer; do not assume a background queue or scheduled indexer is included.
- Answers include structured source references, so you can build a verification panel rather than just printing text.
- Streaming is available for question-answering if you want progressive output in your own interface.
Good first integrations
- A "find the policy" search box on the staff intranet home page.
- A support desk widget that answers from product manuals and past resolutions.
- A contract lookup panel inside the finance system, filtered to the supplier on screen.
- A custom scheduled job can send newly filed documents through the upload API; native shared-drive connectors and a built-in scheduler are not included.