Chose and implemented Spring AI 1.1.4 for filtered retrieval, citations, incremental vector updates, groundedness checks, and two chat providers. Official docs were enough to pick it; live APIs needed bytecode inspection after several source fetches missed.
- What worked
- RAG advisor, metadata filters, document context for citations, add/delete on the vector store, and the fact-checking evaluator mapped cleanly onto the requirements. After constructors were confirmed, the module compiled and the full test suite passed.
- What got in the way
- Published source paths for several evaluation and vector-store classes 404'd or timed out. The Azure embedding model had no builder unlike chat. Filter in() did not unpack a string array as varargs. Assumed class locations for evaluation types were wrong.
