Optional EvalPort export for Dealer.SearchResult — portable retrieval-eval interchange

[adhabnr-ux]

Hi RAGFlow team — I maintain EvalPort, an open JSON-Schema spec for portable LLM eval data (TestCase/Suite/Grader/Result/ResultSet/GraderResult).

I saw discussion #1259 asking for RAG evaluation support, and read rag/nlp/search.py directly. Dealer.SearchResult (total, ids, query_vector, field, highlight, aggregation, keywords, group_docs) is the retrieval result every chunk search in RAGFlow already produces — each id in ids with its field[id] metadata/score is close to an EvalPort GraderResult (chunk_id → grader target, score → score), and a batch of queries run against it is the shape of an EvalPort ResultSet.

I'd propose a standalone ragflow-openeval-adapter package (to_openeval(search_result, query) / from_openeval(suite)), tested against a real Dealer.search() call, not a mock — purely optional, zero footprint on this repo. It would let someone diff RAGFlow's retrieval quality against another framework's graders on identical inputs, which seems relevant to what #1259 was asking for.

Happy to build it as a package in EvalPort's own adapters/ directory, or discuss a small in-repo module if that's preferred.

Spec: https://github.com/adhabnr-ux/evalport/blob/main/SPEC.md

— Sahi, independent contributor (not affiliated with RAGFlow/InfiniFlow)