atlas-rag-evaluation
Skill2 starsChanged 5 months agoApache-2.0View on GitHub
Evaluate RAG quality: run Precision@10 and MRR against a bundle, interpret the scores, and compare CI-built vs locally-built bundles. Use when assessing bundle quality, comparing backends, or investigating a regressed search result.
In this repository16 other files it gives its agents.
- architecture-decision-recordsSkill
- atlas-add-toolSkill
- atlas-build-bundleSkill
- atlas-ci-releaseSkill
- atlas-optimization-patternsSkill
- embedding-strategiesSkill
- hybrid-search-implementationSkill
- mcp-builderSkill
- python-code-styleSkill
- python-observabilitySkill
- python-resilienceSkill
- python-testing-patternsSkill
- python-type-safetySkill
- readme-maintenanceSkill
- skill-creatorSkill
- uv-package-managerSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 35
If the eval script is already registered as a console script.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
When to use
- User asks to "evaluate RAG quality", "run the evaluation script", "check search quality"
- User is comparing MLX vs ONNX+CPU embedding quality
- User wants to verify that a CI-built bundle meets the quality bar
- User suspects a regression in search relevance after a change
Workflow
1. Select a bundle
List available bundles:
ls data/bundles/Pick the one you want to evaluate.
2. Run the evaluation
# Point BUNDLE_DIR at the bundle and run
BUNDLE_DIR=data/bundles/<timestamp> uv run atlas-evalIf the eval script is already registered as a console script. Otherwise run directly:
BUNDLE_DIR=data/bundles/<timestamp> uv run python atlas/evaluate_rag.py3. Read the results
The output looks like:
Query: "how to create an incident"
P@10: 0.90 MRR: 1.0000
...
Mean P@10 over 10 queries: 0.89
Mean MRR over 10 queries: 0.94Precision@10 (P@10): Fraction of the top-10 results that are relevant. A score of 0.89 means ~9 of 10 returned chunks are on-topic.
Mean Reciprocal Rank (MRR): How high the first relevant result appears. 1.0 means the top result is always relevant; 0.94 means the first relevant result is typically in position 1–2.
4. Benchmark bar
| Metric | Minimum | Target | Excellent |
| Precision@10 | 0.75 | 0.85 | 0.95+ |
| MRR | 0.80 | 0.90 | 0.97+ |
- A score below Minimum means the bundle is not usable — investigate chunk boundaries, model, or embeddings.
- Scores between Minimum and Target may be acceptable for development but should not ship to CI.
- Target or higher is required for a release build.
5. Compare CI vs local
To compare, build two bundles with the same model on CI and locally:
# Local (MLX)
uv run atlas-build --model Xenova/bge-small-en-v1.5
BUNDLE_DIR=data/bundles/<local> uv run atlas-eval
# Download CI bundle from GitHub Releases
gh release download <tag> --pattern '*.tar.zst'
tar --zstd -xf atlas-bundle-*.tar.zst -C /tmp/ci-bundle
BUNDLE_DIR=/tmp/ci-bundle uv run atlas-evalCompare the mean scores. Small differences (<0.02) are expected due to float rounding.
Interpretation
| Delta | Meaning |
| <0.01 | Negligible — bundles are equivalent |
| 0.01–0.03 | Minor — likely from int8 quantisation noise |
| 0.03–0.05 | Moderate — investigate chunk boundary differences |
| >0.05 | Significant — model or data divergence |
References
README.md§11 — Validate RAG qualityatlas/evaluate_rag.py— The evaluation script- MTEB leaderboard for model-specific expected scores
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

