An assistant over your own data

The assistant is a conversational surface over the platform. Ask a question about current activity and it searches the platform's own material; upload a document you would not otherwise put in a shared tool and it answers over that as well, combining what it finds in both into a single reply with the sources it used shown alongside.

It is built for work that cannot leave the network. Every model call — both the tool-calling loop and the final synthesis — goes to a locally hosted endpoint, and no external model path exists in the deployment, which means the same assistant can be used on material that has no business leaving the building.

• Chat sessions — conversations kept as sessions with their own history.
• Uploads — documents parsed, chunked and embedded so they can be retrieved later.
• Combined answers — one reply synthesised from both the platform's material and your own.
• Cited sources — the articles and chunks a reply used, shown with the reply.

Uploaded files are parsed according to their type, split into chunks and embedded, then stored with their vectors beside the text. Retrieval is scoped in the database to the person who owns those chunks, so isolation does not depend on the model's judgement or on a filter the model is asked to respect.

The assistant decides what to consult from the question, calls the tools it needs, and streams the answer back as it is written, with the tool calls visible as they happen rather than only in the finished result. A long search therefore looks like a long search, not like a stall.

• Platform search — keyword search across the platform's own articles, with the material it used cited back to the source.
• Own-data search — similarity search over the chunks belonging to the person asking, enforced by owner scope in the query.
• Combined synthesis — one reply that draws on both bodies of material instead of forcing a choice between them.
• Streaming — the answer arrives progressively, and the steps taken to reach it are visible.

The assistant can only call tools that are registered — a small, fixed set covering search, retrieval, enrichment, ingestion and export. There is no shell, no configuration surface and no credential handler in the registry, so the class of action that cannot be expressed cannot be attempted by any prompt.

Anything that changes data or triggers work is gated twice: by the caller's role, and by an approval request put to the user inside the conversation. An approval is single-use, scoped to its owner and valid for a short window, and a second decision on the same request is rejected.

Read-only tools run without ceremony; anything that writes or starts work waits for a person to approve it in the conversation.

Every tool attempt is recorded with the user, the session, the arguments and the outcome — approved, denied, succeeded or failed — and the audit write fails closed. If the record cannot be written, the call does not run, which keeps the log a description of what happened rather than what was intended.

Sessions, uploads and chunks are filtered to their owner in the query itself, and a request for a resource belonging to someone else returns not-found rather than an error that would confirm it exists. That difference matters when the resources are private documents.

• Owner scope — every query filters on the calling user in SQL, never by instruction to the model.
• Role gates — tools that cause work are limited to the roles that should be able to cause it.
• Approval record — the decision, the arguments and the outcome are all kept.
• Stored uploads — files are kept per user under sanitised names, never by the original filename.

The practical effect is that a question spanning a private document and the live picture can be asked in one place, on infrastructure the operator controls, without the document being copied into a third-party service to make it searchable. That matters most for material that is useful precisely because it has not been published anywhere.

And because the same platform material is what the assistant searches, a claim in a reply can be traced back to the article or the data behind it. The assistant is useful in proportion to how checkable it is.

The assistant answers from what it can retrieve, so a question about material it has not ingested gets an honest miss rather than a confident guess. It is a retrieval and synthesis tool, not an authority on the data it reads.

Model, embedding model and upload limits are configuration, not hidden behaviour. They are set by whoever runs the deployment, and the audit trail records what the assistant actually did when it was asked.