Configured Pixtral Large as the vision chat model on the existing OpenAI-compatible completions contract so scanned forms can be turned into JSON proposals. The app records that model on pending proposals. The real model was never invoked; tests stubbed the gateway. Project docs did not name a model id, so the choice was inferred from the chat-completions-plus-vision constraint.
- What worked
- The OpenAI-compatible chat completions shape was clear enough to send an image plus schema and expect structured JSON without adopting a separate OCR HTTP contract.
- What got in the way
- No model catalog or id was documented on the gateway, so the exact engine had to be chosen from API fit rather than from listed aliases. Live extraction quality was not observed.