How semantic search finds things you can't grep

grep is wonderful and it will never find the note you're thinking of unless you remember the exact words you wrote. Type "staging database password" and a note titled "RDS creds, pre-prod" stays hidden, because not one of those characters matches.

Semantic search fixes that by matching meaning instead of characters. The trick is to stop comparing text and start comparing numbers.

Text becomes vectors

An embedding model turns a piece of text into a list of numbers (a vector) that represents where that text sits in "meaning space". Things that mean similar things land near each other, even with no words in common.

"staging database password"  ->  [0.02, -0.41, 0.18, ... ]   (768 numbers)
"RDS creds, pre-prod"         ->  [0.03, -0.39, 0.21, ... ]   (close by!)
"recipe for banana bread"     ->  [-0.55, 0.12, -0.30, ... ]  (far away)

The first two end up close together because the model learned, from huge amounts of text, that "staging" and "pre-prod" are related, and that "creds" and "password" are near synonyms.

Closeness = relevance

To search, you embed the query the same way, then find the stored vectors nearest to it. "Nearest" is usually cosine similarity, the angle between two vectors:

• angle near 0° → very similar
• angle near 90° → unrelated

score = cos(θ) between query vector and each stored vector
return the top N highest scores

That's the whole idea. No keyword has to match; the geometry does the work.

Why it's fast at scale

Comparing your query against every vector is fine for hundreds of notes but slow for millions. A vector index (HNSW, IVF, and friends) trades a tiny bit of accuracy for a huge speedup, so lookups stay quick as the collection grows.

inode uses sqlite-vec by default so the index lives in the same local file as your data, and can switch to pgvector when you want a bigger, shared store.

The catch

Semantic search is fuzzy on purpose, so it can surface something almost-right with confidence. The fix most tools use, including inode, is to hand the top matches to an LLM that re-reads them and answers from what's actually there, rather than trusting the vector score alone. Vectors find the candidates; the model picks the answer.