Mintplex-Labs
[FEAT]: Attach documents to messages
What would you like to see?
Currently, only administrators and managers can upload documents, and then those documents must be added to a workspace and accessed through RAG / pinning.
It would be good if users could upload attachments to messages. Attachments would be processed / OCRed and added to the context of the current conversation, but not added to the RAG system / made available to other users. This is a very common usage pattern, supported by ChatGPT, Claude, etc.
@petrusali https://github.com/Mintplex-Labs/anything-llm/issues/2405 The issue was quite vague and ambiguous. If this issue is what you were asking we need a more verbose use-case and description as above.
We have to be pretty strict on pruning issues since a ton of people use GH issues for help/support or things get out of hand very fast for us.
I haven't found a workaround; we've been advising users to copy-paste document content into the message box in the mean time.
What I am thinking is to Pin the docs which are currently dropped in the chat and unpin all other docs by doing some modifications
I've run into this issue today - a client was very confused about why the file they uploaded wasn't fully referenced by the AI modal being used.
I was surprised to see file uploads actually get embedded in the workspace, rather than just being "pinned" to the thread. It's an odd UX choice, as now a private file from a private thread has been embedded into the workspace and is available to all users of that workspace.
Related issues:
- https://github.com/Mintplex-Labs/anything-llm/issues/3347
- https://github.com/Mintplex-Labs/anything-llm/issues/1543
- https://github.com/Mintplex-Labs/anything-llm/issues/3287
- https://github.com/Mintplex-Labs/anything-llm/issues/3137
Just adding some feedback to this one. I've got two fairly decent sized companies requesting this feature (each with about 30 users on the Docker version).
I'm getting inundated with questions about file uploads, and they tend to fall into two categories:
- I can't upload files (they have the default permission, so aren't allowed to upload files to a workspace)
- The AI isn't able to read the document (they expect the whole document to be kept in the context window, rather than a RAG lookup based on prompts)
I'm concerned that without the ability to simply upload a file to a chat thread, users will simply stop using AnythingLLM and revert back to big AI.
This is a real sticking point at the moment and could ultimately mean that people don't use the platform at all.
We actually switched to Claude Enterprise a few months ago; the models were significantly better than anything we could run on-premise (enough to justify the privacy risks) and the UI was less flaky and more feature rich than any open-source UI we tried.