Where each method fails

Neither one is the good one

The two methods fail in opposite directions, and the failures are predictable enough to plan around.

Semantic search cannot find an identifier. Ask for invoice INV-2024-0142 and the embedding has no notion of it — the string carries no topic, so the vector is noise and the ranking is arbitrary. Exact tokens are precisely what a keyword index is for.

Keyword search cannot find a synonym. Ask about "car insurance" of a corpus that says "vehicle cover" and BM25 scores zero: no shared terms, no match, however obviously related the two are.

This is why serious retrieval systems run both. Not for extra recall in general — for covering each other's specific, known blind spot.

Your turn: write keyword_hits(query_terms, documents) returning the indexes of the documents containing all the query terms. This is the exact-match half, and its strictness is the point.

def keyword_hits(query_terms, documents):
    """Indexes of documents containing every query term."""
    return []