Is RAG_USE_FULL_CONTEXT All or Nothing?

[rwmehta]

If I set the config var for the RAG api "RAG_USE_FULL_CONTEXT" to true, then does that mean its true for everything regardless of file size? Is there some way to have it so it only triggers for files under a certain size?

I want it to be true because I'm dealing with a lot of CSVs and Excel files so the default of top_k=4 doesn't work, but if it is a large file (above 128k tokens or close to it) I'm worried it'll fill up the context way too fast.

Thanks!

[rwmehta]

Also as a follow up question, if this is set to true then is the embedding/actual rag functionality disabled? And how does this interact with resend files?

[ala85]

Yes, check createContextHandlers.js where if enabled it will fetch whole content fro database (as there is a text column together with the embedding) instead of query k neighbor from embedding database

[rwmehta]

Yes, check createContextHandlers.js where if enabled it will fetch whole content fro database (as there is a text column together with the embedding) instead of query k neighbor from embedding database

I see, so it will apply to every file regardless of their size.

Do you by chance know if there is a way to avoid embedding the file in the first place if I want to use full context? I'd like to be able to run this without an embedding model needed.

[ala85]

Not aware, as you need to store the file somewhere. Ofc you can customize the code by reading file in memory (if not too big) or store it in your own table. Another case is to upload a file as text, which I think relies on OCR Mistral at the moment (https://www.librechat.ai/docs/features/ocr) and available for agents only, but @danny-avila can answer better on that

[stdweird]

see also #6605