RAG Facts Check

A modular, high-performance system for verifying RAG-generated answers against their source documents using claim extraction, targeted evidence retrieval, and per-claim verification.

Designed as a drop-in replacement for legacy verification backends (such as Halloumi), it integrates seamlessly with the EEA Chatbot frontend (volto-eea-chatbot).

---

Architecture Overview

flowchart TD
    subgraph Input["Input Data"]
        ANS["RAG Answer\n(prose or markdown tables)"]
        DOCS["Source Documents\n(HTML, PDFs, Web briefings)"]
    end

    subgraph Pipeline["RAG Facts Check Pipeline"]
        CHUNK["split_answer_into_chunks()\n(Preserves table headers & boundaries)"]
        EXTRACT["ClaimExtractor\n(Multi-chunk extraction + span dedup)"]
        VERIFY["ClaimVerifier\n(Batch verification with KV cache reuse)"]
        SPANS["Robust Evidence Span Matcher\n(Whitespace, line-wrap & quote tolerant)"]
        AGG["RAGFactsChecker._aggregate\n(0-10 Answer Score & Report)"]
    end

    subgraph Output["Output Formats"]
        REPORT["CheckReport\n(Groundedness, dimensions, flags)"]
        HA["Halloumi Adapter\n(/halloumi/generate -> claims + segments)"]
    end

    ANS --> CHUNK
    CHUNK --> EXTRACT
    DOCS --> VERIFY
    EXTRACT -->|"Claims [1..N]"| VERIFY
    VERIFY -->|"Evidence quotes + Doc Index"| SPANS
    DOCS --> SPANS
    SPANS -->|"Character offsets {start, end}"| AGG
    AGG --> REPORT
    AGG --> HA

Loading

---

End-to-End Integration Flow

The service acts as the verification backbone for the EEA Chatbot in Plone/Volto:

sequenceDiagram
    autonumber
    actor User
    participant Volto as Volto Frontend (AIMessage.tsx)
    participant Plone as Plone Backend / Chatbot API
    participant Proxy as Express Proxy (/_ha/generate)
    participant Checker as RAG Facts Check (:8000)
    participant LLM as Local LLM Gateway (:4002)

    User->>Volto: Ask question ("What is the EU doing to combat climate change?")
    Volto->>Plone: Send chat message
    Plone->>LLM: Tool search + answer generation
    Plone-->>Volto: Stream response (answer + citations [1], [2] + source documents)
    Volto->>Volto: Filter context sources (qualityCheckContext = 'citations')
    Volto->>Proxy: POST /_ha/generate { answer, sources }
    Proxy->>Checker: POST /halloumi/generate
    Checker->>Checker: split_answer_into_chunks(answer)
    loop For each chunk
        Checker->>LLM: Extract atomic claims
        LLM-->>Checker: Structured claims with verbatim fragments
    end
    Checker->>Checker: Deduplicate identical spans & re-index
    Checker->>LLM: Verify claims in batches against source documents
    LLM-->>Checker: Verdicts, confidence, verbatim evidence quotes, doc_index
    Checker->>Checker: Match evidence spans across newlines & format segments
    Checker-->>Proxy: { answer_score: 8.5, claims: [...], segments: {...} }
    Proxy-->>Volto: JSON response
    Volto-->>User: Render inline claim highlights, score pill & citation modals

Loading

---

Key Capabilities

1. Chunked Claim Extraction for Long Answers & Tables

• The Challenge: Long answers (>2,000 characters), especially markdown tables with 8+ rows and concluding sections, exceed the LLM output token limit when extracted in a single pass.
• The Solution: split_answer_into_chunks() parses markdown tables, keeps row integrity, replicates table headers into every split chunk so column context is retained, and merges short headers/summaries.
• All extracted claims are aggregated, deduplicated by verbatim text spans in the answer, and re-indexed.

2. Robust Multi-Strategy Evidence Span Matching

• The Challenge: Scraped documents (PDFs, web pages) have hard linebreaks (\n), irregular whitespace, and formatting artifacts. LLMs frequently add quotes or ellipses when citing evidence.
• The Solution: find_evidence_span_in_doc() runs a resilient 4-stage matching algorithm:
  1. Quote & Ellipsis Stripping: Cleans surrounding straight (", '), smart (“, ”), and bracketed quotes.
  2. Exact Match: Instant offset matching when verbatim text matches.
  3. Whitespace-Flexible Regex (\s+): Matches single-line quotes against documents with line breaks.
  4. Punctuation & Word-Boundary ([\W_]+): Matches text across hyphenation and punctuation variations.
  5. Fuzzy Sequence Alignment: Fallback to SequenceMatcher for minor paraphrasing.

3. Answer Quality Score (0–10)

• Aggregates per-claim verdicts into an interpretable 0–10 score:
  • Groundedness Base: Supported claims count for 1.0, unverified claims count for 0.4.
  • Citation Penalty: Up to 30% reduction if supported claims fail to cite source text.
  • Contradiction Penalty: Harsh reduction when claims contradict source documents.

4. Batch Verification with KV-Cache Prefix Reuse

• When batch_size > 1 (default: 20), claims are grouped into batches.
• Static documents are placed before claims in the prompt, allowing backends like llama.cpp and vLLM to reuse pre-computed KV caches for massive speedups.

---

Quick Start

Installation

# Clone the repository
git clone git@github.com:eea/rag-facts-check.git
cd rag-facts-check

# Create virtual environment and install dependencies
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[test,dev,server]"

Configuration (.env)

Create a .env file in the root directory:

LLM_API_BASE=http://localhost:4002/v1
LLM_API_KEY=sk-litellm-master-key
LLM_MODEL=gemma
LLM_TEMPERATURE=0.1

Running the Web Service

# Development server with auto-reload
make serve

# Or run directly via uvicorn
uvicorn rag_facts_check.server:app --reload --host 0.0.0.0 --port 8000

Production Docker

docker build -t rag-fact-check .
docker run -p 8000:8000 -e LLM_API_BASE=http://host.docker.internal:4002/v1 rag-fact-check

---

API Endpoints

POST /halloumi/generate

Drop-in replacement for the Halloumi middleware used by volto-eea-chatbot.

• Request:

  {
    "answer": "The European Climate Law sets a binding net-zero target by 2050.",
    "sources": [
      {
        "title": "European Climate Law",
        "text": "The European Climate Law sets a binding target to achieve climate neutrality by 2050.",
        "source_type": "web",
        "link": "https://eur-lex.europa.eu/..."
      }
    ],
    "batch_size": 20
  }

• Response:

  {
    "answer_score": 9.5,
    "claims": [
      {
        "claimString": "The European Climate Law sets a binding net-zero target by 2050.",
        "startOffset": 0,
        "endOffset": 64,
        "segmentIds": ["0"],
        "score": 1.0,
        "rationale": "Document confirms this explicitly.",
        "skipped": false
      }
    ],
    "segments": {
      "0": {
        "id": 0,
        "startOffset": 0,
        "endOffset": 89
      }
    }
  }

POST /check

Full RAG fact-checking endpoint returning detailed analytical dimensions, hallucination flags, and per-claim verdicts.

GET /health

Returns service status and version ({"status": "ok", "version": "0.2.0"}).

---

Testing

The project includes an extensive test suite covering chunking, claim deduplication, span alignment, retrieval, and FastAPI endpoints.

# Run pytest
pytest

# Run with test coverage
pytest --cov=rag_facts_check

# Run linter
ruff check tests/ rag_facts_check/

• Current Status: 190 tests passing (100% pass rate).
• Core Coverage: 87% on checker.py, 88% on spans.py, 95% on retriever.py, 100% on models.py.

---

Documentation Index

Full documentation is available in docs/:

• Project Overview — Core goals, challenges, and architecture
• System Architecture — Module layout and component responsibilities
• Data Flow & Lifecycle — Step-by-step pipeline, sequence diagrams, and span mapping
• Answer Quality Score — Scoring formula, weights, and calibration
• Web Service Guide — Endpoints, request schemas, and integration examples
• Testing Guide — Test suite layout, fixtures, and coverage

---

License

MIT License. Developed for the European Environment Agency (EEA).