Askgrep

Yes for that task. On 9 SWE-bench Lite instances BM25 put the gold file in the top ten every time while askgrep managed five of nine, because issue text is a similarity query, not a does-this-code-do-X predicate. Use lexical search for issue-to-file, askgrep when you can state the predicate.

The repo reports two wins, one loss and one void row from evals.py, which runs offline from committed data with no key. The loss is the SWE-bench comparison above, and the void row is flagged as not a benchmark because its answer key came from the same model family as a subject.

No. --dry-run counts chunks and prices the sweep without a key, and every number on the page was measured with Jev at $0.042 per million input tokens with output free. You can also use any OpenAI-compatible endpoint, including a local model, though endpoints without logprobs return only 0.0 or 1.0.

Almost always phrasing. On the same 120 chunks, abstract phrasing like panics instead of returning an error gave recall 0.07 while the literal contains a call to .unwrap() gave 0.71. Phrase the question as something the code does and rewrite before concluding the tool cannot see it.

cargo install --git https://github.com/fajarhide/askgrep, then export TYPESAFE_API_KEY. Release binaries exist for macOS on Apple silicon or Intel and for Linux on musl. Flags: -t/--threshold defaults to 0.8, -j/--jobs sets concurrency at 16 by default, --dry-run counts and prices a sweep without calling anything, --json emits one object per hit for piping into jq, and --no-cache re-asks every chunk. Scores from different backends never share a cache.

Answers are cached on disk by content, so re-running a question over unchanged files costs nothing and the sweep, fix, sweep again loop is free after the first pass. That is what makes the price on the page checkable: the demo tree returns two hits in 11 chunks for 3,910 tokens and $0.0002, and a second identical sweep spends nothing. The README's framing is that a full sweep of its own repository covers 819 functions for $0.0049, against $0.12 for Haiku and $0.58 for Opus on the same tokens.

askgrep needs one number per chunk and each chunk carries different code, so there is no shared document to amortise. That is the opposite case from TypeSafe's published batching result, where many questions about one shared document in a single call is 12.2 times cheaper, 10 times faster and changes no answer. Both can be true: batch when the questions read the same text, split when each question reads its own.