Allow external providers to set a custom instruction for LLMs to understand the text

[kyteinsky]

Describe the feature you'd like to request

For analytics or emails, the LLM just looks at the content and has no idea what it's looking at. Adding hints there everytime is tedious for the user.

Describe the solution you'd like

The external provider should supply that for each of its content providers, which can then be sent to CBB and to an LLM.

Describe alternatives you've considered

/

[marcelklehr]

Don't we inject the type of document in the llm prompt?

[kyteinsky]

ah right we have that "type" field. But for files, it's the mimetype, they share the same field.
we may have to separate and annotate them before sending for inference.
ref: https://github.com/nextcloud/context_chat_backend/blob/ab2adcd831a5fdb0f4d9a2351c202f5ce6f613ac/context_chat_backend/chain/ingest/injest.py#L82
in here maybe: https://github.com/nextcloud/context_chat_backend/blob/ab2adcd831a5fdb0f4d9a2351c202f5ce6f613ac/context_chat_backend/chain/context.py#L35

[marcelklehr]

But for files, it's the mimetype, they share the same field.

Wut? Mh. That was not the goal of the type field 🤔
Can we send the mimetype to the backend separately and keep the type as File?

[kyteinsky]

oops, the meaning was lost in variable names it seems :)

Can we send the mimetype to the backend separately and keep the type as File?

the DB in CCB would need a migration to add a column and make the modifications.

[kyteinsky]

now there is a dedicated and explicit field for provider description: nextcloud/server#63459