Working guide: AFFiNE + LiteLLM + Ollama via Responses→Chat bridge (FIXED)

[axcode07]

AFFiNE Copilot streams to /v1/responses (Responses API) and expects a specific SSE prelude + a stable item_id. LiteLLM/Ollama expose OpenAI Chat Completions (/v1/chat/completions) with a different stream shape—causing the classic “text part … not found” error.

What this is:
A tiny Node adapter that accepts POST /v1/responses, translates inputs into OpenAI-style messages[], calls LiteLLM chat (stream or non-stream), and re-emits proper Responses-API events. /v1/models and /v1/embeddings are passed through unchanged. No patching AFFiNE, no forking LiteLLM.

Why it’s simple:

• Normalizes AFFiNE content parts → plain string messages

• Emits the required Responses prelude (response.output_item.added, response.content_part.added, …)

• Streams deltas cleanly; stable item_id

Works with LiteLLM + Ollama out of the box

Check out repo: https://github.com/axcode07/affine_ai_helper

Quick start

• Run the adapter container (Dockerfile included).

• In your proxy (e.g. Caddy), route only /v1/responses* to the adapter; everything else → LiteLLM.

In AFFiNE → Copilot (OpenAI), keep baseURL = https:///v1 and your API key.

Feedback and forks are welcomed!

[iacademiahub]

It worked like magic! The show WILL GO ON thanks to @axcode07

[Sawtaytoes]

This seems like a great solution. I wanna confirm a few things though:

1. Can I use Nginx-Proxy-Manager instead of Caddy?
2. Sounds like the crux is LiteLLM? Are you running in that in Docker? It's not in the TrueNAS Apps, so I'll have to manually configure and maintain it.
3. How did you configure your settings in AFFiNE? I changed all my models and the API lookups to be Ollama, but that's not gonna work (as you said), so I need to switch it to your Node.js app (another Docker container), which points to LiteLLM (Docker container), and then I configure LiteLLM to target Ollama?
4. And it looks like you do need to modify openai.ts based on your README :/.