RAG without vectors

The question

For corpora under ~10k documents, do you actually need a vector DB.

Setup

3,200 markdown documents. Same retrieval questions. Same downstream LLM. Compared Postgres tsvector with BM25-ish ranking against pgvector with text-embedding-3-large. Eval set was 50 questions.

Result

tsvector top-3 recall: 92%. pgvector top-3 recall: 94%. Answer quality from the downstream LLM was indistinguishable between the two.

Verdict

For small corpora full-text search is fine and one less moving part. Switch to vectors when recall starts to hurt.