ragscore

Generate QA datasets & evaluate RAG systems in 2 commands

🔒 Privacy-First • ⚡ Lightning Fast • 🤖 Any LLM • 🏠 Local or Cloud • 🌍 Multilingual

English | 中文 | 日本語 | Deutsch

⚡ 2-Line RAG Evaluation

# Step 1: Generate QA pairs from your docs
ragscore generate docs/

# Step 2: Evaluate your RAG system
ragscore evaluate http://localhost:8000/query

That's it. Get accuracy scores and incorrect QA pairs instantly.

============================================================
✅ EXCELLENT: 85/100 correct (85.0%)
Average Score: 4.20/5.0
============================================================

❌ 15 Incorrect Pairs:

  1. Q: "What is RAG?"
     Score: 2/5 - Factually incorrect

  2. Q: "How does retrieval work?"
     Score: 3/5 - Incomplete answer

🚀 Quick Start

Install

pip install ragscore              # Core (works with Ollama)
pip install "ragscore[openai]"    # + OpenAI support
pip install "ragscore[notebook]"  # + Jupyter/Colab support
pip install "ragscore[all]"       # + All providers

Already installed? Keep up to date — new versions add features like failure diagnosis and retrieved context capture:

pip install --upgrade ragscore

Option 1: Python API (Notebook-Friendly)

Perfect for Jupyter, Colab, and rapid iteration. Get instant visualizations.

from ragscore import quick_test

# 1. Audit your RAG in one line
result = quick_test(
    endpoint="http://localhost:8000/query",  # Your RAG API
    docs="docs/",                            # Your documents
    n=10,                                    # Number of test questions
)

# 1b. Tailored QA — target specific audiences
result = quick_test(
    endpoint="http://localhost:8000/query",
    docs="docs/",
    audience="developers",                   # Who asks the questions?
    purpose="api-integration",               # What's the document for?
)

# 2. See the report
result.plot()

# 3. Inspect failures
bad_rows = result.df[result.df['score'] < 3]
display(bad_rows[['question', 'rag_answer', 'reason']])

Rich Object API:

• result.accuracy - Accuracy score
• result.df - Pandas DataFrame of all results
• result.plot() - 3-panel visualization (4-panel with detailed=True)
• result.corrections - List of items to fix

Option 2: CLI (Production)

Generate QA Pairs

# Set API key (or use local Ollama - no key needed!)
export OPENAI_API_KEY="sk-..."

# Generate from any document
ragscore generate paper.pdf
ragscore generate docs/*.pdf --concurrency 10

# Tailored QA generation — target specific audiences
ragscore generate docs/ --audience developers --purpose faq
ragscore generate docs/ --audience customers --purpose "pre-sales"
ragscore generate docs/ --audience "compliance auditors" --purpose "security audit"

Evaluate Your RAG

# Point to your RAG endpoint
ragscore evaluate http://localhost:8000/query

# Custom options
ragscore evaluate http://api/ask --model gpt-4o --output results.json

🔬 Detailed Multi-Metric Evaluation

Go beyond a single score. Add detailed=True to get 5 diagnostic dimensions per answer — in the same single LLM call.

result = quick_test(
    endpoint=my_rag,
    docs="docs/",
    n=10,
    detailed=True,  # ⭐ Enable multi-metric evaluation
)

# Inspect per-question metrics
display(result.df[[
    "question", "score", "correctness", "completeness",
    "relevance", "conciseness", "faithfulness"
]])

# Radar chart + 4-panel visualization
result.plot()

==================================================
✅ PASSED: 9/10 correct (90%)
Average Score: 4.3/5.0
Threshold: 70%
──────────────────────────────────────────────────
  Correctness: 4.5/5.0
  Completeness: 4.2/5.0
  Relevance: 4.8/5.0
  Conciseness: 4.1/5.0
  Faithfulness: 4.6/5.0
==================================================

Metric	What it measures	Scale
Correctness	Semantic match to golden answer	5 = fully correct
Completeness	Covers all key points	5 = fully covered
Relevance	Addresses the question asked	5 = perfectly on-topic
Conciseness	Focused, no filler	5 = concise and precise
Faithfulness	No fabricated claims	5 = fully faithful

CLI:

ragscore evaluate http://localhost:8000/query --detailed

🔍 Failure Diagnosis (--diagnose)

When answers fail, --diagnose tells you why — retriever miss, generator hallucination, incomplete answer, or wrong interpretation:

ragscore evaluate http://localhost:8000/query --diagnose

🔍 Failure Diagnosis:
  Retriever Miss: 3 (42.9%)
  Generator Hallucination: 2 (28.6%)
  Incomplete Answer: 1 (14.3%)
  Wrong Interpretation: 1 (14.3%)

Uses the support_span already generated with each QA pair to give the judge grounding context. Combine with --detailed for full diagnostics:

ragscore evaluate http://localhost:8000/query --diagnose --detailed -o results.json

Category	Meaning
Retriever Miss	RAG didn't retrieve the chunk containing the evidence
Generator Hallucination	Retrieved correctly but fabricated information
Incomplete Answer	Retrieved correctly but answer is partial
Wrong Interpretation	Retrieved correctly but misunderstood the content

📓 Full demo notebook — build a mini RAG and test it with detailed metrics.

🎯 Audience & Purpose demo — generate tailored QA for developers, customers, auditors, and more.

🏠 Ollama local demo — 100% private RAG evaluation with no API keys.

🏠 100% Private with Local LLMs

# Use Ollama - no API keys, no cloud, 100% private
ollama pull llama3.1
ragscore generate confidential_docs/*.pdf
ragscore evaluate http://localhost:8000/query

Perfect for: Healthcare 🏥 • Legal ⚖️ • Finance 🏦 • Research 🔬

Ollama Model Recommendations

RAGScore generates complex structured QA pairs (question + answer + rationale + support span) in JSON format. This requires models with strong instruction-following and JSON output capabilities.

Model	Size	Min RAM	QA Quality	Recommended
llama3.1:70b	40GB	48GB VRAM	Excellent	GPU server (A100, L40)
qwen2.5:32b	18GB	24GB VRAM	Excellent	GPU server (A10, L20)
llama3.1:8b	4.7GB	8GB VRAM	Good	Best local choice
qwen2.5:7b	4.4GB	8GB VRAM	Good	Good local alternative
mistral:7b	4.1GB	8GB VRAM	Good	Good local alternative
llama3.2:3b	2.0GB	4GB RAM	Fair	CPU-only / testing
qwen2.5:1.5b	1.0GB	2GB RAM	Poor	Not recommended

Minimum recommended: 8B+ models. Smaller models (1.5B–3B) produce lower quality support spans and may timeout on longer chunks.

Ollama Performance Guide

# Recommended: 8B model with concurrency 2 for local machines
ollama pull llama3.1:8b
ragscore generate docs/ --provider ollama --model llama3.1:8b

# GPU server (A10/L20): larger model with higher concurrency
ollama pull qwen2.5:32b
ragscore generate docs/ --provider ollama --model qwen2.5:32b --concurrency 5

Expected performance (28 chunks, 5 QA pairs per chunk):

Hardware	Model	Time	Concurrency
MacBook (CPU)	llama3.2:3b	~45 min	2
MacBook (CPU)	llama3.1:8b	~25 min	2
A10 (24GB)	llama3.1:8b	~3–5 min	5
L20/L40 (48GB)	qwen2.5:32b	~3–5 min	5
OpenAI API	gpt-4o-mini	~2 min	10

RAGScore auto-reduces concurrency to 2 for local Ollama to avoid GPU/CPU contention.

🔌 Supported LLMs

Provider	Setup	Notes
Ollama	ollama serve	Local, free, private
OpenAI	export OPENAI_API_KEY="sk-..."	Best quality
Anthropic	export ANTHROPIC_API_KEY="..."	Long context
DashScope	export DASHSCOPE_API_KEY="..."	Qwen models
vLLM	export LLM_BASE_URL="..."	Production-grade
Any OpenAI-compatible	export LLM_BASE_URL="..."	Groq, Together, etc.

📊 Output Formats

Generated QA Pairs (output/generated_qas.jsonl)

{
  "id": "abc123",
  "question": "What is RAG?",
  "answer": "RAG (Retrieval-Augmented Generation) combines...",
  "rationale": "This is explicitly stated in the introduction...",
  "support_span": "RAG systems retrieve relevant documents...",
  "difficulty": "medium",
  "source_path": "docs/rag_intro.pdf"
}

Evaluation Results (--output results.json)

{
  "summary": {
    "total": 100,
    "correct": 85,
    "incorrect": 15,
    "accuracy": 0.85,
    "avg_score": 4.2
  },
  "incorrect_pairs": [
    {
      "question": "What is RAG?",
      "golden_answer": "RAG combines retrieval with generation...",
      "rag_answer": "RAG is a database system.",
      "score": 2,
      "reason": "Factually incorrect - RAG is not a database"
    }
  ]
}

🧪 Python API

from ragscore import run_pipeline, run_evaluation

# Generate QA pairs
run_pipeline(paths=["docs/"], concurrency=10)

# Generate tailored QA pairs for specific audiences
run_pipeline(
    paths=["docs/"],
    audience="support engineers",
    purpose="fine-tuning a support chatbot",
)

# Evaluate RAG
results = run_evaluation(
    endpoint="http://localhost:8000/query",
    model="gpt-4o",  # LLM for judging
)
print(f"Accuracy: {results.accuracy:.1%}")

🤖 AI Agent Integration

RAGScore is designed for AI agents and automation:

# Structured CLI with predictable output
ragscore generate docs/ --concurrency 5
ragscore evaluate http://api/query --output results.json

# Exit codes: 0 = success, 1 = error
# JSON output for programmatic parsing

CLI Reference:

Command	Description
ragscore generate <paths>	Generate QA pairs from documents
ragscore generate <paths> --audience <who>	Tailored QA for specific audience
ragscore generate <paths> --purpose <why>	Focus QA on document purpose
ragscore evaluate <endpoint>	Evaluate RAG against golden QAs
ragscore evaluate <endpoint> --detailed	Multi-metric evaluation
ragscore evaluate <endpoint> --diagnose	Failure root-cause classification
ragscore --help	Show all commands and options
ragscore generate --help	Show generate options
ragscore evaluate --help	Show evaluate options

⚙️ Configuration

Zero config required. Optional environment variables:

export RAGSCORE_CHUNK_SIZE=512          # Chunk size for documents
export RAGSCORE_QUESTIONS_PER_CHUNK=5   # QAs per chunk
export RAGSCORE_WORK_DIR=/path/to/dir   # Working directory

🔐 Privacy & Security

Data	Cloud LLM	Local LLM
Documents	✅ Local	✅ Local
Text chunks	⚠️ Sent to LLM	✅ Local
Generated QAs	✅ Local	✅ Local
Evaluation results	✅ Local	✅ Local

Compliance: GDPR ✅ • HIPAA ✅ (with local LLMs) • SOC 2 ✅

🧪 Development

git clone https://github.com/HZYAI/RagScore.git
cd RagScore
pip install -e ".[dev,all]"
pytest

📡 Telemetry

RAGScore collects telemetry only in MCP server mode (ragscore serve). Standard CLI and Python API usage do not send telemetry.

We collect limited anonymous operational metrics to understand feature usage and improve reliability. No document content, prompts, QA text, model outputs, API keys, endpoint URLs, or file paths are collected.

Collected in MCP mode:

• MCP tool invoked
• LLM provider and model name
• ragscore version, Python version, OS type
• Success/failure status
• Random anonymous installation ID

Opt out:

export RAGSCORE_NO_TELEMETRY=1

�� Links

• GitHub • PyPI • Issues • Discussions

⭐ Star us on GitHub if RAGScore helps you!
Made with ❤️ for the RAG community