Semantic Search

Semantic search is a retrieval method that ranks documents by the similarity between their meaning and a query's meaning — computed as the distance between embedding vectors — rather than by matching literal keywords.

A keyword search engine looks for the same characters the user typed. Ask it "how do I stop my laptop from overheating" and it will miss a support article titled "reducing thermal throttling," because the two share almost no words. Semantic search closes that gap by comparing meaning instead of strings.

How it works

Every document (or document chunk) is embedded once, offline, into a vector, and stored in an index. At query time, the user's question is embedded with the same model, and the system finds the stored vectors closest to it by cosine similarity — typically returning the top k as ranked results. No keyword ever has to match; the ranking is purely geometric.

The quality of a semantic search system rests on three levers: the embedding model (how well it captures domain-specific meaning), the chunking strategy feeding it (badly sized chunks blur or starve context), and the index serving the nearest-neighbor lookup (exact for small collections, approximate like HNSW at scale). Weakness in any one of the three shows up as worse rankings, and it is easy to misdiagnose which lever actually broke.

Why it matters

Semantic search is the retrieval half of nearly every "search my documents" or "find things like this" product feature, and it is the R in retrieval-augmented generation. Its known weakness is the mirror of keyword search's weakness: an embedding model has little reason to distinguish rare identifiers, product SKUs, or exact error codes — it was trained to cluster meaning, not memorize strings — so semantic search alone often misses queries that contain one. That's precisely the gap hybrid search is built to close by running keyword and semantic search together.

A worked example

A gear-shop knowledge base holds three chunks: "Our 3-season tents handle wind and light rain down to 20°F," "Return any unused item within 60 days for a full refund," and "Down sleeping bags rated to -20°F for winter expeditions." A user searches "camping gear for freezing temperatures." A keyword match finds nothing — no chunk contains "freezing" or "camping gear" verbatim. Semantic search embeds the query and compares: the sleeping bag chunk scores highest (its content is about cold-weather gear even without shared words), the tent chunk scores moderately (cold-adjacent but less extreme), and the returns chunk scores lowest (unrelated topic entirely). The ranking reflects meaning, and the top result is one a keyword engine would never surface.

How Miatz teaches it

Learners build a keyword index and a semantic index over the identical corpus, then run a batch of paraphrased queries — questions that deliberately avoid the source document's exact words — against both, and watch keyword search's recall collapse while semantic search holds. The exercise runs immediately before hybrid search, so the two methods' complementary failure modes are visible before the fix is taught.