Benchmarks

Recall is not free: it costs tokens on every prompt. These are the measured numbers from this repository, including the cases where it does not pay off.

The standing cost

Recall costs about 180 tokens a prompt, and is skipped entirely on 28% of prompts because the answer is already in context.

The refund

Asking who defines and calls closeObservations by hand means a grep, then opening three store adapters to work out which hits are definitions. One explain call answers it already classified: 2.1x cheaper, measured across eight real symbols in this repo.

  every prompt costs      180 tokens of recall
  every lookup refunds    597 tokens   (1,115 by hand - 518 with)

  so one lookup pays for 3.3 prompts of recall

Where it loses

A symbol with two or three references is cheaper to grep than to look up, because the graph answer carries structure the question did not need. The full write-up lists every symbol measured, including those.

Run it yourself

git clone https://github.com/Avijit07x/claude-db
cd claude-db && npm install && npm run build

npm run bench:ab      # with and without, per symbol
npm run bench:tokens  # what recall costs per prompt

Both run against whatever repository you point them at, so the interesting number is the one from your own codebase rather than from this one.