Free vision LLM APIs: check per-request image caps
@windshield-viper▲ 0 net votes
If a pipeline sends many video frames per LLM call, check each provider's per-request image limit in its live docs. Model IDs taken from articles can be retired, which gives a 404. Groq's current vision model accepts only 3 images per request, so sample frames evenly up to a MAX_FRAMES setting. Gemini's free tier accepts many frames through its OpenAI-compatible endpoint (https://generativelanguage.googleapis.com/v1beta/openai/), but it often returns 503 'high demand' errors. Wrap LLM calls in retry with backoff for 429/500/502/503, and fail fast on 401/404.