What is semantic search?
Search that matches what a passage means, not the words it happens to use. It is the step that makes a document library answerable.
Semantic search finds passages by meaning rather than by matching words. Your question and every passage in a library are converted into vectors, long number lists that encode what the text is about, and the search returns the passages whose vectors sit closest to the question's. A passage answering your question in entirely different wording still ranks. That is the retrieval step behind answering from a document library, and it runs on your machine.
Match the meaning, not the wording.
Keyword search asks whether the document contains your words. Semantic search asks whether the document is about your meaning. The conversion happens through embeddings: a small model reads a passage and produces a vector, and passages about similar things end up with similar vectors, whatever words they use. Ask about "income inequality" and a passage that never uses the phrase but discusses wealth gaps still surfaces. The vectors are compared by distance, which is arithmetic, not comprehension, and that cheapness is the point.
Where it sits in the loop.
In a grounded research tool, semantic search is the selector that decides what the model gets to read. A question arrives, the library is searched, and the few most relevant passages are placed in front of the model to answer from. That retrieval step is what separates a grounded answer from a remembered one. Pure vector matching misses exact terms and names, and pure keyword matching misses paraphrase, so serious local tools run both methods and merge the results. The whole chain, embedding the question, searching the library, running the model, executes locally on modest hardware.
Approximate by nature.
Matching by meaning means matching approximately. Near-miss passages, ones that discuss a similar topic without answering the question, can outrank the true source, and no score fixes that on its own. This is why retrieval is the front of the pipeline and not the end of it: search narrows the field to candidates, and a verification pass checks what the answer claims against the passage it cites. The search finds where an answer might live; only the check says whether it does. The other half of retrieval quality sits before the search even runs: what a retrieval query is covers how your question becomes the text the index sees.