v0.3.0 — code and document search

sgrep now searches code and documents with Minish Lab’s static embeddings. It chooses Potion Base 8M for prose-only candidates and Potion Code 16M v2 for code or mixed candidates; --model code and --model text let you choose explicitly.

The README includes a Context.dev benchmark graphic, links to alternatives, model credits, and the full methodology. The graphic presents the existing v0.2.0 code-search results; it does not claim a general-text benchmark.

Install with cargo install --git https://github.com/mrmps/sgrep --locked, or extract the attached macOS Apple Silicon executable onto your PATH. Ripgrep is required. Source builds need Rust 1.90 or later. Each selected model downloads about 32–34 MB on first use and then works offline.

All 31 real-model E2E checks passed, including manual model selection and offline use. All 60 code-benchmark ranked result lists were identical to v0.2.0. Fresh text-model download and separate cache behavior were also verified. Only macOS ARM was exercised.

The attached PNG and SVG are ready to share; verification.json contains the check receipts.

v0.2.0 — native Rust search

Native Rust CLI: ripgrep + Potion Code static embeddings, no source index or daemon.

On 100 RepoQA Python queries, fresh-process median latency fell from 397 ms to 74 ms (5.4x); p95 fell from 834 ms to 129 ms. All 100 ranked result lists matched the Python version exactly. M5 Max / 48 GB, cached model files, filesystem caches not cleared.

Install from source with cargo install --git https://github.com/mrmps/sgrep --locked (Rust 1.90+). Remove the old pip/uv installation first. The attached macOS arm64 binary avoids compilation; ripgrep must still be on PATH. First matching search downloads about 34 MB of model files, then searches work offline.

Only macOS arm64 was exercised. The included binary passed 22 real-model CLI checks, including fresh edits, ignore rules, nested Python functions, offline operation, and closed output pipes.

Comparison on a fixed 60-query RepoQA retrieval sample: 12 repositories, six languages. Latency is a separate sequential 12-query panel on the same M5 Max.

Tool	Median	Correct file first	Half function within 8k tokens	Full function within 8k
sgrep	71 ms	53.3%	76.7%	66.7%
ripgrep	23 ms	0%	15.0%	1.7%
CK	226 ms	43.3%	76.7%	75.0%
dzhng/jevgrep + TypeSafe	10,183 ms	75.0%	60.0%	60.0%

Jevgrep has the best first-file accuracy; its Java answers often provide locations without source excerpts, which score zero on source coverage. It uses the native retrieval core, TypeSafe jev-1.13.0, default rate budgeting, no answer cache, and a resident adapter excluding CLI startup. These are complete search times, not individual model-request latency.

CK's index setup took 6m 51s and 54.7 MB, excluded from search time. CK returns complete functions more often than sgrep. Ripgrep is literal OR matching in path order, not agent-generated grep queries. These small source snapshots total 12 MB; this is a 10% sample and does not establish full-benchmark or large-monorepo performance.

Detailed report · Per-query scores · Protocol, recorded outputs and runner source