Used the core framework plus OpenAI, Anthropic, and embedding adapters for chunking, provider abstraction, citations, and faithfulness evaluation. The APIs were capable and inspectable, though model-specific adapter behavior required checking package source.
- What worked
- The framework supplied useful abstractions for embeddings, generation, citation handling, and groundedness evaluation while allowing Django to retain authorization and persistence control.
- What got in the way
- Current model behavior was not always obvious from adapter documentation, so wheel source was inspected to understand temperature handling and model recognition.