Providers and models¶
Last reviewed: 2026-08-14
The Course uses LiteLLM so provider-specific request plumbing does not obscure RAG. Model availability, quotas, privacy terms, and prices change. Verify them before deployment.
Learning paths¶
| Path | Recommendation | Trade-off |
|---|---|---|
| Free generation | Google Gemini free tier | Quotas vary; review how free-tier data may be used. |
| Inexpensive paid generation | OpenAI GPT-5.6 Luna | Requires billing; suitable for repeatable Course runs. |
| Fast free experiments | Groq | Model availability and free limits can change. |
| Model comparison | OpenRouter | One API for many models; free capacity is not a production SLA. |
| Local embeddings | sentence-transformers/all-MiniLM-L6-v2 |
Stable and CPU-friendly, but not a universal production embedding model. |
Hugging Face Inference Providers gives free accounts only a small monthly credit, so BuildRAG does not present it as a substantial free generation path.
Keep provider keys in environment variables, pin the LiteLLM version, and record the exact model with every evaluation run.