hawkes.aiKey ideas
- The one idea — Semantic search finds text by meaning, not keywords: embed the query and every document into vectors (meaning is a direction), then retrieve the documents whose vectors point most nearly the same way as the query's.
- How it is shown — "How do I reset my password" retrieving a page titled "account recovery steps" — no shared words, but the meaning-vectors point the same way.
- The trap to avoid — Picturing keyword matching under the hood — it compares meaning-directions, not shared words.
- What it sets up — Searching purely by meaning can fumble exact names and codes.
Search 'reset my password' and the system finds 'account recovery steps' — zero words in common. It wasn't matching words at all. It was matching directions.
The one idea
Semantic search finds text by meaning, not keywords: embed the query and every document into vectors (meaning is a direction), then retrieve the documents whose vectors point most nearly the same way as the query's.
Everything in RAG rides on one step: finding the right passages. And good search doesn't look for matching words. It searches by meaning. Ask how do I reset my password, and it finds a page titled account recovery steps — no shared words at all. Recall an idea from earlier: meaning is a direction. An embedding turns any text into a vector — a point in meaning-space — so texts that mean similar things point in similar directions, and unrelated ones point far apart. That embedding is the engine under semantic search. Here's how it works. Ahead of time, you embed every chunk of your knowledge base into vectors. When a query arrives, you embed it too.
How it works — the demo
"How do I reset my password" retrieving a page titled "account recovery steps" — no shared words, but the meaning-vectors point the same way.
Then you find the chunks whose vectors point most nearly the same way as the query's. Closest in meaning-space wins. So relevance becomes geometry. The closeness between two directions measures how related two texts are — a tiny gap for near-synonyms, a wide one for unrelated ideas. Retrieve the query's nearest neighbors, and you've found the passages that mean what the question means. That's why it's powerful. It matches paraphrases, synonyms, related ideas. Search automobile and it finds car. Search a concept, and it surfaces text that never used your exact terms — text a plain keyword search would walk right past. The trap: picturing keyword matching under the hood.
The trap to avoid
Picturing keyword matching under the hood — it compares meaning-directions, not shared words.
Why it matters — and what’s next
Searching purely by meaning can fumble exact names and codes.
It isn't. Nothing scans for shared words; it compares meaning-directions. That's its strength — and, as we'll see, occasionally its weakness, when the exact word is what matters. So semantic search finds text by meaning: embed everything, embed the query, grab the nearest neighbors. Retrieval by concept, not word. But searching purely by meaning can fumble the exactly-literal — a product code, a rare name. So do keywords still matter? Next: keywords aren't dead.
How can search find a document sharing not one word with your query? Because meaning, to a machine, is a direction — and two directions can point the same way.
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
Everything in RAG rides on one step: finding the right passages. And good search doesn't look for matching words. It searches by meaning. Ask how do I reset my password, and it finds a page titled account recovery steps — no shared words at all.
Recall an idea from earlier: meaning is a direction. An embedding turns any text into a vector — a point in meaning-space — so texts that mean similar things point in similar directions, and unrelated ones point far apart. That embedding is the engine under semantic search.
Here's how it works. Ahead of time, you embed every chunk of your knowledge base into vectors. When a query arrives, you embed it too. Then you find the chunks whose vectors point most nearly the same way as the query's. Closest in meaning-space wins.
So relevance becomes geometry. The closeness between two directions measures how related two texts are — a tiny gap for near-synonyms, a wide one for unrelated ideas. Retrieve the query's nearest neighbors, and you've found the passages that mean what the question means.
That's why it's powerful. It matches paraphrases, synonyms, related ideas. Search automobile and it finds car. Search a concept, and it surfaces text that never used your exact terms — text a plain keyword search would walk right past.
The trap: picturing keyword matching under the hood. It isn't. Nothing scans for shared words; it compares meaning-directions. That's its strength — and, as we'll see, occasionally its weakness, when the exact word is what matters.
So semantic search finds text by meaning: embed everything, embed the query, grab the nearest neighbors. Retrieval by concept, not word. But searching purely by meaning can fumble the exactly-literal — a product code, a rare name. So do keywords still matter? Next: keywords aren't dead.