eyuansu71/agent_retrieval_bench

Agent Retrieval Bench

Agent Retrieval Bench is a curated code-retrieval benchmark for coding agents. It evaluates whether a retrieval system can find the repository files an agent would need for real software-engineering workflows.

The current public release is Agent Retrieval Bench V1: 225 manually curated samples across three tasks.

Project page: https://agent-retrieval-bench.github.io/

Task	Samples	What is retrieved
code2test	106	Tests related to an implementation change or PR intent
comment2context	51	Extra context files needed to act on a review comment
trace2code	68	Root-cause source files for a reproduced failure trace
Total	225

V1 is frozen. Future data expansion will use a new version rather than mutating benchmark/v1.

Licensing note: benchmark metadata, reports, documentation, and evaluator code are MIT licensed; redistributed corpus chunks remain governed by their upstream repository licenses. See the evaluator repository for LICENSE, DATA_LICENSE.md, and CITATION.cff.

Files

Recommended download path:

releases/v1/agent_retrieval_bench_v1.tar.zst
releases/v1/agent_retrieval_bench_v1.tar.zst.sha256

The release bundle extracts to:

benchmark/v1/        # JSONL samples and manifest
corpus/v1/           # chunked repository corpora at base commits
eval/v1/             # baseline summaries/details
reports/v1/          # leaderboard, status, analysis

The older v1_code_review track is still available for historical comparisons, but full V1 is the primary benchmark.

Quickstart

Install the evaluator from GitHub:

pip install "git+https://github.com/eyuansu62/agent-retrieval-bench.git"

Download and extract V1 with the evaluator CLI:

arb download-benchmark --version v1 --local-dir data --force

This downloads the release bundle from Hugging Face, verifies the checksum, and extracts benchmark/v1, corpus/v1, eval/v1, and reports/v1.

Manual troubleshooting path:

hf download eyuansu71/agent_retrieval_bench \
  --repo-type dataset \
  --local-dir data \
  --include "releases/v1/**"

cd data
shasum -a 256 -c releases/v1/agent_retrieval_bench_v1.tar.zst.sha256
zstd -dc releases/v1/agent_retrieval_bench_v1.tar.zst | tar -xf - -C .
cd ..

Validate samples:

arb validate data/benchmark/v1/*.jsonl

Run the lexical baseline:

arb eval-baseline \
  --derived data/benchmark/v1 \
  --corpus data/corpus/v1 \
  --out data/eval/v1/lexical_summary.json \
  --details data/eval/v1/lexical_details.jsonl \
  --no-keep-list

Run the deterministic vectorless RepoMap baseline:

arb eval-repomap \
  --derived data/benchmark/v1 \
  --corpus data/corpus/v1 \
  --out data/eval/v1/repomap_summary.json \
  --details data/eval/v1/repomap_details.jsonl \
  --candidate-filter all_files \
  --no-keep-list

Generate a leaderboard from available *_summary.json files:

arb report-models \
  --eval-dir data/eval/v1 \
  --out data/reports/v1/model_leaderboard.md \
  --json-out data/reports/v1/model_leaderboard.json

Project page: agent-retrieval-bench.github.io. Source code, evaluator commands, and implementation details are in the GitHub repository: eyuansu62/agent-retrieval-bench.

Current Leaderboard

Primary ranking metric: overall MRR. Candidate set: all_files.

Task	Model	Samples	Recall@5	Recall@10	Recall@20	MRR	Gold@8k
overall	Qwen3-Embedding-4B	225	0.2883	0.4033	0.5828	0.2455	0.2542
overall	aider-style-repomap	225	0.3089	0.4705	0.6299	0.2227	0.0704
overall	jina-code-embeddings-0.5b	225	0.2230	0.3133	0.4492	0.1883	0.1556
overall	lexical	225	0.1970	0.3267	0.4874	0.1450	0.0785
code2test	Qwen3-Embedding-4B	106	0.4610	0.5777	0.7230	0.3225	0.3887
code2test	jina-code-embeddings-0.5b	106	0.2610	0.3868	0.5305	0.2033	0.2060
code2test	aider-style-repomap	106	0.2752	0.4107	0.5808	0.1975	0.0739
code2test	lexical	106	0.0676	0.1399	0.2469	0.0663	0.0299
comment2context	jina-code-embeddings-0.5b	51	0.3431	0.4020	0.5261	0.3282	0.2386
comment2context	Qwen3-Embedding-4B	51	0.2451	0.3856	0.5621	0.3113	0.2451
comment2context	aider-style-repomap	51	0.2353	0.3137	0.4967	0.2053	0.0588
comment2context	lexical	51	0.1928	0.3595	0.5752	0.1739	0.0784
trace2code	aider-style-repomap	68	0.4167	0.6814	0.8064	0.2750	0.0735
trace2code	lexical	68	0.4020	0.5931	0.7966	0.2458	0.1544
trace2code	Qwen3-Embedding-4B	68	0.0515	0.1446	0.3799	0.0760	0.0515
trace2code	jina-code-embeddings-0.5b	68	0.0735	0.1324	0.2647	0.0598	0.0147

Key observation: Qwen is strongest overall and on code2test; Jina is strongest on comment2context; RepoMap is strongest on trace2code. Weak embedding scores on trace2code are an intended benchmark signal: that task rewards structured failure-log and repo-graph retrieval, not only semantic code similarity.

Evaluation Semantics

• Every sample is evaluated against the repository at the recorded base commit; fixed code is not indexed.
• comment2context treats the commented file as given context and scores only additional required context files.
• trace2code uses audited root-cause source files as primary gold; related tests are not primary gold.
• gold_coverage@8k measures whether gold files appear within an 8k-character retrieval budget.

Versioning

• Primary benchmark: benchmark/v1
• Earlier code-review-only track: benchmark/v1_code_review
• Legacy smoke benchmark: benchmark/v0_2

Planned V1.1 work is targeted expansion of comment2context and trace2code; code2test is not expanded by default.

Citation

There is no paper yet. For now, cite the project page and dataset:

@misc{agent_retrieval_bench_v1,
  title        = {Agent Retrieval Bench V1},
  author       = {{Agent Retrieval Bench contributors}},
  year         = {2026},
  howpublished = {\url{https://agent-retrieval-bench.github.io/}},
  note         = {Dataset: \url{https://huggingface.co/datasets/eyuansu71/agent_retrieval_bench}}
}

Machine-readable citation metadata is available in the evaluator repository: CITATION.cff.