← Notes

Between not indexing and indexing, there is the textual middle

The cheap path is not a choice between nothing and a full vector pipeline: classic keyword search combined with embeddings is the recommended arrangement, and its textual half charges nothing per query.

The indexing discussion is usually framed as binary: either the corpus has no search, or you build the whole vector pipeline. Anthropic's contextual retrieval material describes the arrangement the industry uses, and it is hybrid: "By leveraging both BM25 and embedding models, traditional RAG systems can provide more comprehensive and accurate results, balancing precise term matching with broader semantic understanding".

The interesting part for anyone with a budget constraint is that the two halves have very different prices. Term search — exact word matching, which is what BM25 does — runs inside the database, calling no model, and therefore with no cost per query. The embeddings are what cost, at ingestion and at every reprocessing.

That suggests an order rather than a choice. Start with the half that does not charge: textual search over the notes, measuring whether what comes back answers the real questions that arrive. If it does, the vector pipeline becomes an optional improvement instead of a prerequisite. If it does not, there is now a concrete case — the question textual search got wrong — to justify the cost of the other half.

The gain from that order is not technical, it is budgetary: it defers the spend until there is evidence it buys something.