ἵστωρ Search Follow the build

What is an embedding?

The quiet conversion that makes meaning-level search possible. Every grounded answer over a library depends on it happening once, well.

An embedding turns a passage of text into a list of numbers that captures what it means. Passages about similar things get similar lists, so a question can be matched to sources by sense rather than by exact wording. This is the step that lets retrieval find a passage even when it shares no keywords with your question, and it runs fine on your own machine.

Text as coordinates.

A computer cannot compare two paragraphs directly, but it can compare two lists of numbers without effort. An embedding model reads a passage and outputs one such list, often several hundred values long, chosen so that meaning maps to position: two passages about the same finding land near each other even if one says efficacy and the other says how well it worked. Nothing magical sits in the numbers. They are the residue of a model trained on enormous amounts of text, and their nearness is a usable stand-in for sameness of subject.

one library, two ways to reach ita paper: efficacya thesis: how well it workeda paper: adverse eventsa note: drug safetythe question: drug safetykeywords reach the wording, not the subjectnearness reaches across the wording one library, two ways to reach ita paper: efficacya thesis: how well it workeda paper: adverse eventsa note: drug safetythe question: drug safetykeywords reach the wording, not the subjectnearness reaches across the wording
The same library, drawn as positions. Two passages about one finding sit nearest each other while saying different words; keywords reach the note that shares the wording, and nearness reaches the subject across it. The positions are the argument, not a model's own coordinates.

Why search needs them.

Keyword search has an honest strength and an honest limit. It never misses an exact term, and it cannot see a paraphrase. Ask about drug safety while a paper says adverse events, and keywords return nothing while the meaning sits right there; search built on this trick is called semantic search, and it is the other half of the pairing. This is why retrieval-augmented generation systems run both searches together: keywords for the terms you actually used, embeddings for the ones you did not. The store that files those embeddings and answers the near-neighbor question is a vector database, which on a personal scale is smaller and simpler than the name sounds. In Istor the two searches are blended, so a quoted phrase still finds its source and a reworded concept still finds its passage.

Where the conversion happens.

Embedding models are small by language-model standards, small enough to run on a laptop CPU without anyone noticing. That fact carries a privacy consequence: the conversion that makes your documents searchable does not require sending them anywhere. In a local tool the embeddings are computed on your machine and stored in a local database, which is part of what keeps a local RAG loop private. When you evaluate any tool that searches documents, the question to ask is simply where that step runs.

Related