hawkes.aiKey ideas
- The one idea — Semantic search blurs exact terms (names, codes, acronyms) while keyword search nails them; the best systems use hybrid (both) plus a reranker that re-scores the top candidates for true relevance.
- How it is shown — A query for exact code "XR-500" that semantic search fumbles but keyword search nails; hybrid retrieves broadly, then a reranker floats the truly-best passage to the top.
- The trap to avoid — Assuming semantic search replaced keywords — they're complementary; keyword for exactness, semantic for meaning, rerank for precision.
- What it sets up — Embedding millions of chunks needs somewhere to store and search them fast.
Type an exact product code into semantic search and watch it return vaguely related junk. Meaning-search blurs precise strings. That's why keyword search never actually died.
The one idea
Semantic search blurs exact terms (names, codes, acronyms) while keyword search nails them; the best systems use hybrid (both) plus a reranker that re-scores the top candidates for true relevance.
Semantic search feels like magic — until you search for an exact product code, XR-500, and it hands back vaguely related junk. Meaning-search blurs precise tokens. It's brilliant at concepts and clumsy at exact strings. That's why keyword search never died. Think about what each is good at. Semantic search catches paraphrases and ideas, but can miss an exact name, an error code, a rare acronym. Keyword search — the classic word-matching kind — nails those exact terms, but misses anything phrased differently. So the best systems don't choose. They go hybrid: run both searches, semantic and keyword, and combine the results.
How it works — the demo
A query for exact code "XR-500" that semantic search fumbles but keyword search nails; hybrid retrieves broadly, then a reranker floats the truly-best passage to the top.
One brings meaning, the other precision. Together they catch far more of what's relevant than either alone — the concept matches and the exact-string matches, pooled into one haul. Then a second stage: reranking. Hybrid gives you a broad shortlist. A reranker — a slower, sharper model — reads each candidate against the query and re-scores it for true relevance, floating the genuinely best passages to the top. So the pattern is: retrieve broadly, then rerank precisely. Cast a wide net with hybrid search so the right passage is somewhere in the pile, then let the reranker sort it to the front. Recall wide, precision late — cheaper and more reliable than nailing the best match in one shot. The trap: believing semantic search replaced keywords.
The trap to avoid
Assuming semantic search replaced keywords — they're complementary; keyword for exactness, semantic for meaning, rerank for precision.
Why it matters — and what’s next
Embedding millions of chunks needs somewhere to store and search them fast.
It didn't. They're complementary — keyword for exactness, semantic for meaning, a reranker for precision. And the deeper point: real RAG quality is won here, in retrieval, not by a fancier model. So keywords aren't dead: hybrid plus a reranker beats naive semantic-only, often by a lot. You can now find the right text reliably. But embedding millions of chunks and searching them in milliseconds needs somewhere to store them. Next: what a vector database is for.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
Semantic search feels like magic — until you search for an exact product code, XR-500, and it hands back vaguely related junk. Meaning-search blurs precise tokens. It's brilliant at concepts and clumsy at exact strings. That's why keyword search never died.
Think about what each is good at. Semantic search catches paraphrases and ideas, but can miss an exact name, an error code, a rare acronym. Keyword search — the classic word-matching kind — nails those exact terms, but misses anything phrased differently.
So the best systems don't choose. They go hybrid: run both searches, semantic and keyword, and combine the results. One brings meaning, the other precision. Together they catch far more of what's relevant than either alone — the concept matches and the exact-string matches, pooled into one haul.
Then a second stage: reranking. Hybrid gives you a broad shortlist. A reranker — a slower, sharper model — reads each candidate against the query and re-scores it for true relevance, floating the genuinely best passages to the top.
So the pattern is: retrieve broadly, then rerank precisely. Cast a wide net with hybrid search so the right passage is somewhere in the pile, then let the reranker sort it to the front. Recall wide, precision late — cheaper and more reliable than nailing the best match in one shot.
The trap: believing semantic search replaced keywords. It didn't. They're complementary — keyword for exactness, semantic for meaning, a reranker for precision. And the deeper point: real RAG quality is won here, in retrieval, not by a fancier model.
So keywords aren't dead: hybrid plus a reranker beats naive semantic-only, often by a lot. You can now find the right text reliably. But embedding millions of chunks and searching them in milliseconds needs somewhere to store them. Next: what a vector database is for.