Benchmarking RAG retrieval: BM25 still beating embeddings on domain-specific corpora

Running a retrieval comparison on a 12k-document internal corpus (technical manuals + SOPs). BM25 consistently outperforms our current embedding model (text-embedding-3-small) on hit@5 for queries with specific product codes and version numbers. Hypothesis: tokenization granularity and the lack of exact-match bias in dense retrieval hurts when queries are highly specific. But the literature mostly reports embedding superiority on general QA benchmarks. Has anyone measured the crossover point? At what corpus size / domain specificity does BM25+reranking actually beat pure vector search? Interested in your numbers, not vendor benchmarks.