OpenPipe support

[enricoros]

Support the OpenPipe service, to the following features levels:

• [x] recording of chats by filtering them through the OpenPipe service
• [x] creating fine-tunes for the user (this will be done through the OpenPipe App)
• [ ] listing of fine-tunes (will need an API for that)
• [ ] chatting with finetunes (would be supported)

Stretch goals may or may not be:

• [ ] upload of existing chats to OpenPipe, to seed the database/finetuning data
• [ ] not having OpenPipe as a separate service, but add it as an option to the OpenAI service
• [ ] supporting Anthropic and other models - this may be done by OpenPipe through the proxy service (routing the request and translating the params) - or by us, by having separate provider-transformers to upgrade the requests

Why doing this:

• enable people to get more value out of their chats by finetuning a model which may answer according to their preferences
• dedicated models shall be cheaper and replace: in-context learning, ragging, heavy prompting

This ticket is to track the support for the service. For more information see OpenPipe or our configuration dialog.

[enricoros]

Initial implementation landed in the big-agi-2 branch (warning; the branch is in a dangerous and experimental state - only for big-agi devs).

Screenshots/recordings:

https://github.com/user-attachments/assets/591cde9a-17ad-4ef2-9f50-1fdbefb82327

[enricoros]

Update - OpenPipe now provides the models list api - https://docs.openpipe.ai/api-reference/get-listModels. Need to integrate it.

[enricoros]

Update: the API is now listing all the OpenPipe Fine-Tunes, in addition to 4 OpenAI models that are pre-configured.

[enricoros]

As of 8b88d4afe72bdb53b92d38cfeb4a9c2e832541af - recording, model listing, and chatting work:

[enricoros]

Working:

• [x] recording, with custom tagging (to OpenAI models)
• [x] listing of fine-tune models
• [x] chatting with fine-tunes

Future work:

• [ ] make it work as an option to any OpenAI model (instead of a standalone service)
• [ ] batch upload of user chats
• [ ] pre-filter/warn for 8k context limits
• [ ] decode/demarshall output in case of fine-tune learned chat encoding
• [ ] surface this functionality as a well polished feature to the users (pmf to be tuned)