rag-eval

Evaluate your RAG pipeline quality using Ragas metrics (faithfulness, answer relevancy, context precision).

Security Scan

The skill is internally consistent with its stated purpose (RAG evaluation using Ragas) — it asks for the expected LLM keys, runs ragas via Python, and its instructions match the included scripts — but review and safe installation practices (virtualenv, inspect code) are recommended before running.

Purpose & Capability

Name/description (RAG evaluation with Ragas) aligns with code and instructions. Declared required binaries (python3, pip), optional env vars (OPENAI/ANTHROPIC/RAGAS_LLM), and the included scripts (run_eval.py, batch_eval.py, setup.sh) are all appropriate for performing LLM-judged RAG evals.

Instruction Scope

SKILL.md instructs the agent to accept question/answer/contexts, write a temp JSON file, and call the provided Python scripts; it explicitly warns against shell-injecting user content. The scripts only reference expected files/paths (memory/eval-results) and expected env vars. No instructions request unrelated system data or unrelated credentials.

Install Mechanism

No registry install spec is provided; the included scripts/setup.sh installs dependencies via pip from public PyPI packages (ragas, datasets, langchain integrations). This is expected for a Python tool but may modify system Python if a virtualenv isn't used. No downloads from untrusted URLs or URL-shortened installers were found.

Credentials

Requested environment access is limited to LLM-related keys and optional RAGAS_* tuning variables. These are justified by the skill's need to call an LLM judge and (optionally) embeddings. No unrelated secrets or multiple unrelated service credentials are requested.

Persistence & Privilege

The skill does not request always:true and does not modify other skills. It persists evaluation outputs under memory/eval-results (expected for a reporting tool). The setup script may install packages on the host environment but does not request elevated system privileges.

Assessment

This skill appears to do what it claims, but take these precautions before installing or running it: 1) Inspect the included scripts locally (scripts/run_eval.py, scripts/batch_eval.py, scripts/setup.sh) — don't run arbitrary shell scripts without review. 2) Use a Python virtual environment (python -m venv .venv; source .venv/bin/activate) before running setup.sh to avoid global pip installs. 3) Protect your LLM keys — the skill uses OPENAI_API_KEY/ANTHROPIC_API_KEY to call remote LLMs; grant least-privilege keys where possible and monitor usage. 4) The tool writes evaluation files to memory/eval-results in the working directory; verify this location suits your data-retention policies. 5) There is a truncated/possibly buggy section in the provided run_eval.py excerpt (the sample here was truncated) — ensure you have the complete, reviewed script before running explain/advanced features. 6) Expect runtime costs for LLM judge calls. If you need higher assurance, ask for a line-by-line code review or a reproducible test run in an isolated environment.

RAG Eval — Quality Testing for Your RAG Pipeline

Test and monitor your RAG pipeline's output quality.

🛠️ Installation

1. Ask OpenClaw (Recommended)

Tell OpenClaw: "Install the rag-eval skill." The agent will handle the installation and configuration automatically.

2. Manual Installation (CLI)

If you prefer the terminal, run:

clawhub install rag-eval

⚠️ Prerequisites

1. Your OpenClaw must have a RAG system (vector DB + retrieval pipeline). This skill evaluates the output quality of that pipeline — it does not provide RAG functionality itself.
2. At least one LLM API key is required — Ragas uses an LLM as judge internally. Set one of:
   • OPENAI_API_KEY (default, uses GPT-4o)
   • ANTHROPIC_API_KEY (uses Claude Haiku)
   • RAGAS_LLM=ollama/llama3 (for local/offline evaluation)

Setup (first run only)

bash scripts/setup.sh

This installs ragas, datasets, and other dependencies.

Single Response Evaluation

When user asks to evaluate an answer, collect:

1. question — the original user question
2. answer — the LLM output to evaluate
3. contexts — list of text chunks used to generate the answer (retrieved docs)

⚠️ SECURITY: Never interpolate user content directly into shell commands. Write the input to a temp JSON file first, then pipe it to the evaluator:

# Step 1: Write input to a temp file (agent should use the write/edit tool, NOT echo)
# Write this JSON to /tmp/rag-eval-input.json using the file write tool:
# {"question": "...", "answer": "...", "contexts": ["chunk1", "chunk2"]}

# Step 2: Pipe the file to the evaluator
python3 scripts/run_eval.py < /tmp/rag-eval-input.json

# Step 3: Clean up
rm -f /tmp/rag-eval-input.json

Alternatively, use --input-file:

python3 scripts/run_eval.py --input-file /tmp/rag-eval-input.json

Output JSON:

{
  "faithfulness": 0.92,
  "answer_relevancy": 0.87,
  "context_precision": 0.79,
  "overall_score": 0.86,
  "verdict": "PASS",
  "flags": []
}

Post results to user with human-readable summary:

🧪 Eval Results
• Faithfulness: 0.92 ✅ (no hallucination detected)
• Answer Relevancy: 0.87 ✅
• Context Precision: 0.79 ⚠️ (some irrelevant context retrieved)
• Overall: 0.86 — PASS

Save to memory/eval-results/YYYY-MM-DD.jsonl.

Batch Evaluation

For a JSONL dataset file (each line: {"question":..., "answer":..., "contexts":[...]}):

python3 scripts/batch_eval.py --input references/sample_dataset.jsonl --output memory/eval-results/batch-YYYY-MM-DD.json

Score Interpretation

Score	Verdict	Meaning
0.85+	✅ PASS	Production-ready quality
0.70-0.84	⚠️ REVIEW	Needs improvement
< 0.70	❌ FAIL	Significant quality issues

Faithfulness Deep-Dive

If faithfulness < 0.80, run:

python3 scripts/run_eval.py --explain --metric faithfulness

This outputs which sentences in the answer are NOT supported by context.

Notes

• Ragas uses an LLM internally as judge (uses your configured OpenAI/Anthropic key)
• Evaluation costs ~$0.01-0.05 per response depending on length
• For offline use, set RAGAS_LLM=ollama/llama3 in environment