⚡ RAG Harness

Evaluate RAG systems in seconds — fast, deterministic, and without requiring an API key.

Evaluating RAG systems shouldn't require paid APIs, rigid schemas, or complicated evaluation pipelines.

RAG Harness is a lightweight CLI that evaluates your RAG outputs using deterministic metrics, automatically detects common dataset formats, and provides human-readable diagnostics to help you understand why your system succeeds or fails.

Why choose RAG Harness?

• No API key required
• Fully deterministic and reproducible
• Works offline and fast
• Works with common RAG formats automatically
• Great for local development and CI

👉Just give your model output → get evaluation instantly.

✓ Offline
✓ Deterministic
✓ CI-friendly
✓ No API

---

📦 Install

pip install rag-harness

📊 Score Interpretation Guide

RAG Harness uses deterministic scoring, which is typically stricter than LLM-based evaluation.

⚠️ Note

• Scores may appear lower than LLM-based evaluators
• Deterministic scoring is stricter and reproducible
• LLM-based evaluation may give higher scores due to semantic reasoning

👉 Example:

A score of ~0.5 in RAG Harness often corresponds to reasonably good semantic answers, even if not perfectly aligned token-wise.

---

🎥 Demo

---

✨ Features

• ⚡ One-command evaluation
• 🧠 RAGAS-style scoring (no API required)
• 🔍 Works with most JSON / JSONL / CSV RAG outputs
• 📋 Rich CLI summaries
• 💡 Human-readable diagnostics
• 🔥 Top Failed Examples
• 🔄 Auto-detects ground truth
• 📊 Exact Match + F1 + Fuzzy + Context metrics
• ⚔️ Compare multiple RAG systems
• 🧩 Handles messy real-world outputs (LangChain, LlamaIndex, custom)

---

▶️ Quick Start

1. Evaluate predictions only

rag-harness evaluate examples/predictions_a.jsonl

2. Full evaluation (recommended)

rag-harness evaluate examples/predictions_a.jsonl --dataset examples/dataset.jsonl

3. Detailed diagnostics (verbose mode)

rag-harness evaluate examples/predictions_a.jsonl --dataset examples/dataset.jsonl --verbose

Displays:

• 📊 Rich evaluation summary
• 💡 Human-readable diagnostics
• 🔥 Top Failed Examples
• 🟢 Overall verdict for each failed example

4. Compare two RAG systems

rag-harness compare examples/dataset.jsonl examples/predictions_a.jsonl examples/predictions_b.jsonl

---

📊 Example Output

📊 RAG Evaluation Summary

Total             3
F1 Score          0.34
Fuzzy Score       0.60
Context Recall    0.00

🧠 RAGAS Score    0.47

🧠 Insights

• Answers are semantically correct but not precise
• No context detected → retrieval not evaluated

---

📁 Supported Input Formats

RAG Harness automatically detects:

• answer, generated_answer, response
• ground_truth, expected_answer
• contexts, documents, source_documents

Works with:

• LangChain outputs
• LlamaIndex outputs
• Custom RAG pipelines
• Benchmark JSON logs

👉 No strict schema required.

---

🧾 Example Formats

Predictions + Ground Truth

{
  "generated_answer": "...",
  "ground_truth": "...",
  "contexts": ["..."]
}

Predictions only

{
  "answer": "...",
  "contexts": ["..."]
}

⚠️ Note

• Without ground truth → limited evaluation
• With ground truth → full evaluation

---

🧠 Scoring

RAG Harness approximates RAGAS using:

• Exact Match
• F1 Score
• Fuzzy Semantic Matching
• Context Recall

⚠️ Important

• Fully deterministic (no API required)
• Faster and reproducible
• Scores may differ from LLM-based RAGAS

---

⚔️ Compare Systems

rag-harness compare dataset.json pred_a.json pred_b.json

⚔️ RAG Systems Comparison

Metric        A      B
------------------------
F1 Score      0.83   0.45
RAGAS Score   0.72   0.51

🏆 System A wins

---

🚧 Roadmap

Completed

• ✅ Rich CLI output
• ✅ Human-readable diagnostics
• ✅ Overall verdicts
• ✅ Top Failed Examples
• ✅ Compare multiple systems

Coming Soon

• ⏳ Rich metrics dashboard
• ⏳ HTML reports
• ⏳ Dataset Doctor
• ⏳ CSV / Markdown export
• ⏳ Optional LLM evaluation

---

🤝 Contributing

PRs, ideas, and improvements are welcome!

---

👨‍💻 Author

Built by Abhishek Bevinkatti

---

If this helped you evaluate your RAG system, consider starring ⭐ the repo!