Semantic search was supposed to make keyword search obsolete. Then you search for an exact model number, a specific acronym, or a function name — and the fancy vector search confidently returns things that are about the same topic but not the thing you literally typed. It turns out the old keyword search was right about something, and the new semantic search was right about something else. The best retrieval systems stopped choosing between them. Building the GovernAI Research Atlas is where that clicked for me.
Two kinds of "match," and each is blind where the other sees
Keyword search matches strings. It's literal, and that literalness is both its weakness and its superpower. It has no idea that "car" and "automobile" are related — but it will nail an exact match on an obscure identifier, a proper noun, or a rare acronym that a semantic model smooths into something vaguely similar.
Semantic search matches meaning. It understands that "sluggish database" and "slow query performance" are the same intent — but it can drift right past an exact term because, in meaning-space, the precise string isn't special. Each approach is blind exactly where the other is sharp: keyword misses meaning, semantic misses specificity.
Hybrid search: run both, then fuse
The insight is that these aren't competitors — they're complementary sensors, and the move is to run both and combine their results rather than betting on one:
- Semantic for intent, keyword for precision. The vector search casts a wide net by meaning so you find relevant things you didn't phrase exactly; the keyword search guarantees the exact term, ID, or name still surfaces even when the embedding would have blurred it.
- Fuse the two rankings. Each method produces its own ranked list, and you merge them into one — so a result that both agree on rises, and each method covers the other's blind spot instead of overriding it.
- The corpus decides the mix. Research papers, repositories, and governance material carry both fuzzy concepts and exact identifiers, which is exactly the case where neither method alone is enough and the hybrid earns its keep.
The takeaway
The lesson generalizes past search: the newest technique rarely makes the old one worthless — more often it's strong where the old one was weak, and the best systems combine them instead of picking a side. Semantic search didn't kill keyword search; it gave it a partner that covers its blind spots.
Building the Atlas taught me to reach for "both, fused" before "either, alone." The full retrieval architecture is on the project page.
👉 Explore it: www.divyakush.com/projects/governai-research-atlas
Divyakush Punjabi — Full-Stack & AI Systems Engineer
🌐 https://www.divyakush.com · 💼 LinkedIn · 💻 GitHub
Top comments (0)