⚡ Why RAGGround?

Most RAG teams evaluate hallucinations using Ragas or LLM-as-a-judge. In production, this causes 3 critical problems:

1. Unusable Latency: Calling GPT-4 to judge an answer adds 3,000–8,000 ms to every response.
2. Double API Bills: Every token generated requires an extra LLM call to evaluate.
3. No Span Citations: LLM judges give vague scores (0.75) without pinpointing the exact hallucinated sentence or linking claims to documents.

RAGGround solves this with a 2-Tiered Local Engine:

• ⚡ Tier 1 (Sub-1ms): Deterministic token-alignment, sequence matching (LCS), and entity consistency verification.
• 🧠 Tier 2 (~10ms CPU): Quantized INT8 Natural Language Inference (NLI) model executed locally with ONNX Runtime. Zero PyTorch installation required (< 45MB RAM).

🚀 Quickstart

1. Installation

pip install ragground

2. Verify RAG Outputs (Python API)

from ragground import RAGGround

guard = RAGGround()

context = """
Tesla reported Q3 revenue of $25.18 billion, an 8% increase year-over-year.
Operating margin came in at 10.8% with free cash flow of $2.74 billion.
"""

answer = """
Tesla's Q3 revenue rose 8% to $25.18 billion.
Free cash flow reached $2.74 billion.
The company also announced plans to launch humanoid robots by Christmas.
"""

report = guard.verify(context=context, answer=answer)

print(f"Grounded: {report.is_grounded}")          # False
print(f"Score: {report.grounding_score:.2f}")     # 0.67 (2 of 3 verified)
print(f"Latency: {report.latency_ms} ms")         # 11.2 ms

# Inspect flagged hallucinations
for h in report.hallucinations:
    print(f"⚠️ Hallucination: {h.text}")
    print(f"   Status: {h.status.value} (Confidence: {h.confidence:.2f})")

# View augmented answer with exact citations injected
print("\n--- Cited Output ---")
print(report.cited_answer)

📊 Benchmark: RAGGround vs. LLM-as-a-Judge

💻 Command Line Interface (CLI)

Audit any context and answer directly in your terminal:

ragground verify \
  -c "Python 3.12 introduced improved error messages and isolated subinterpreters." \
  -a "Python 3.12 added isolated subinterpreters. It also removed the GIL completely."

Benchmark your local CPU:

ragground benchmark

🏗️ Architecture

[ User Context & Answer ] ──► [ Claim Decomposition ] ──► [ Tier 1: Token & LCS Shortcut (< 1ms) ]
                                                                     │
                                  ┌──────────────────────────────────┴─────────────────────────────────┐
                                  ▼ (High Overlap)                                                     ▼ (Paraphrased / Fuzzy)
                         [ Mark Verified ]                                                [ Tier 2: ONNX NLI Transformer (~10ms) ]
                                  │                                                                    │
                                  └───────────────────────────────┬────────────────────────────────────┘
                                                                  ▼
                                                   [ Citation & GuardReport ]

📚 Documentation

For in-depth guides and references, check the docs/ directory:

• Getting Started
• System Architecture
• API Reference
• Benchmarks
• FastAPI & LangChain Integration Guide
• Custom ONNX Models

📄 License

Apache 2.0 License. See LICENSE for details.