Deterministic — re-run in CI on every commit

Benchmarks you can re-run yourself

Every number on this page comes from one of two sources: a deterministic public benchmark on pinned, real OSS repos that CI re-runs on every commit, and community submissions PR'd into a schema-validated dataset. No synthetic demos, no cherry-picking — every row ships with the exact command that reproduces it.

79–100%gold-file recall on pinned real OSS repos
45–257×fewer tokens than pasting whole files
40pre-registered queries — every one reported
46–66×community-measured reduction on real repos

The public benchmark

Cost and correctness, reported together

A token-reduction number with no correctness number attached is meaningless. The public benchmark scores both at once: does the objectively-correct file land in the context window, and how many tokens did it cost to put it there? Built to be hard to dismiss:

Real, pinned repos

Household-name OSS — requests, click, flask, rich — pinned to exact commit SHAs. Anyone can git checkout and audit; no vendor fixture.

Objective gold, no LLM judge

Each query's gold file is the definition site of a named symbol, verifiable with one rg command. Deterministic — nothing to rig.

Strong baselines, disclosed

Not just naive file dumps: keyword search (ripgrep) and a vector RAG using the same encoder NeuralMind uses. Losses shown, not hidden.

Pre-registered queries

Queries are committed in evals/public/manifest.json before tuning. Every one is reported; re-running yields identical numbers.

requests @ 0e322af877

14 pre-registered queries · tokenizer: tiktoken o200k_base

BackendGold-file recallMean tokens/queryvs full-file
full-file1.0041,729
ripgrep0.7926,5431.6×
embedding-rag1.0060769×
neuralmind0.9693044.9×

Misses 1 of 14 (a two-file cross-file query, one of two gold files retrieved).

click @ 874ca2bc1c

7 pre-registered queries · tokenizer: tiktoken o200k_base

BackendGold-file recallMean tokens/queryvs full-file
full-file1.0078,514
ripgrep0.7945,0591.7×
embedding-rag1.00634123.8×
neuralmind0.7978899.6×

Misses 2 of 7 — the weakest repo in the corpus.

flask @ c12a5d874c

10 pre-registered queries · tokenizer: tiktoken o200k_base

BackendGold-file recallMean tokens/queryvs full-file
full-file1.0059,013
ripgrep0.8526,8912.2×
embedding-rag0.9568785.9×
neuralmind0.9577276.4×

Misses 1 of 10.

rich @ 7f580bdc70

9 pre-registered queries · tokenizer: tiktoken o200k_base

BackendGold-file recallMean tokens/queryvs full-file
full-file1.00232,483
ripgrep1.0043,4375.4×
embedding-rag1.00677343.2×
neuralmind1.00905256.8×

No NeuralMind gold-file misses on this repo.

Read the losses too: across all 4 repos (40 queries), NeuralMind's weighted mean recall is 93.75% (90% found-rate) — click is the weakest repo at 0.79 recall (2 misses of 7); requests and flask each miss 1 query. Where embedding-rag also hits full or near-full recall, it's NeuralMind's own vector-retrieval core in isolation, and it's cheaper on raw tokens — the gap is what the progressive-disclosure assembly layer (project map + symbols + call edges) adds on top of raw top-k chunks. Scoring reuses neuralmind/quality.py verbatim — the same metric code the CI quality gate runs. Full method, baselines, and raw per-query data in the methodology.

# Reproduce it yourself — no trust required. Deterministic:
# your numbers match the tables above to the token.
git clone https://github.com/dfrostar/neuralmind && cd neuralmind
pip install -e . tiktoken
python -m evals.public.run

Community numbers

Real repos, submitted via pull request

Each entry below was produced by running neuralmind benchmark . --json on the submitter's own repository and PR'd into the dataset. CI validates every submission against a schema. NeuralMind never uploads anything automatically — these are deliberate, auditable contributions.

Loading…

Does the ratio hold as repos get bigger?

Each point is one submission — graph node count (a proxy for codebase size) against the average reduction ratio vs loading the whole repo. Hover any point for the project, language, and exact numbers.

Reduction by language

Average reduction ratio per language, with the submission count under each bar — so you can tell "this is solid" apart from "n=1, take with salt".

All submissions

The dataset behind both charts. The verification command for every row is neuralmind benchmark . --json — run it on your own repo and the numbers should reproduce in the same range.

Project Language Nodes Reduction Wakeup tok Query tok Model Submitted by Date Notes
Loading submissions…

Click a column header to sort. Data loads from community-benchmarks.json at page load — no server, no tracking.

Contribute

Add your repo's numbers

The dataset gets stronger with every repo it covers — especially across different languages and sizes. Five minutes, two commands, one PR:

pip install neuralmind
cd /path/to/your-repo
neuralmind build .
# prints your numbers + a ready-to-share JSON blob
neuralmind benchmark . --contribute

Then open a PR adding the blob as one entry in docs/community-benchmarks.json (fields enforced by community-benchmarks.schema.json), or file it via the benchmark submission issue template. CI validates the entry; this dashboard updates automatically once it merges.

5-minute walkthrough →