Files for context with non-local models

[KroetschVialutions]

I've been wondering about the ability to use files for context with models that aren't locally hosted, because a colleague of mine uses GPT a lot for work and provides most of the context via files like documentation and such.

Since it is possible to also connect the app with Azure GPT, he was hoping to be able to use the instance that our company is hosting for work for this, but the upload options are greyed out.
Now I am suspecting that this might be because the model is non-local, so I was wondering if that's the case and if there are plans to add the functionality, that the projects and files can be kept local, even though the model asked about them is not.

Assuming that's technically possible.

[wellkilo]

On current Jan, the attachment control is not disabled merely because the chat model is remote. It is disabled when the selected model is not marked as tool-capable:

disabled={!selectedModel?.capabilities?.includes('tools')}

That is because embedded documents are exposed to the model through Jan's internal RAG tools. The thread route only enables those tools when all three conditions hold: documents exist, Attachments are enabled, and the selected model advertises the tools capability.

For an Azure/OpenAI-compatible deployment, the practical setup on current Jan is:

1. Open the provider/model settings and make sure the Azure model has the Tools capability enabled. Custom/remote capabilities are not always auto-detected.
2. Open Settings → Attachments and enable attachments.
3. Attach documents to the thread or project. For thread files, choose inline or embeddings; project files intentionally use embeddings.
4. Start a new turn/thread after changing model capabilities so the available tool set is refreshed.

The files can be parsed and indexed locally, but “kept local” needs one privacy qualification: when you ask a question, the remote Azure model must receive enough text to answer. In inline mode, parsed document text is inserted into the request; in embeddings mode, Jan's local retrieval layer selects chunks and those retrieved chunks appear in tool results sent to the remote model. The whole original file is not uploaded as an Azure file object by this path, but the relevant content does leave the machine.

Current implementation references: the tool-capability gate in the attachment menu, the three-way RAG availability check, and inline versus embeddings processing.

If the Azure model already has tools enabled and the control is still greyed out, please include the Jan version and the model's capability list; that would indicate a shipped-version or provider-metadata issue rather than an inherent remote-model restriction.