2see

AI-powered fact verification for PDF documents.

Upload a PDF. Extract factual claims. Verify each one against live web evidence. Export a transparent, source-cited report.

---

Why

Reports, whitepapers, marketing documents, and AI-generated content all contain factual claims — market sizes, funding rounds, benchmark scores, regulatory actions, historical events. Most go unchecked.

2see automates the verification pipeline: extract → search → verify → report. Every verdict is grounded in retrieved evidence with source credibility scoring, so you can see exactly why a claim was marked Verified, Inaccurate, False, or Unverifiable.

How It Works

flowchart TD
    A["📄 PDF Upload"] --> B["Text Extraction"]
    B --> C["Claim Extraction"]
    C --> D["Materiality Scoring"]
    D --> E["Evidence Retrieval"]
    E --> F["Verdict Synthesis"]
    F --> G["📊 Report Export"]

    classDef default fill:#2d3436,stroke:#636e72,stroke-width:1.5px,color:#dfe6e9,rx:8
    linkStyle default stroke:#636e72,stroke-width:1.5px

Loading

1. PDF → Text — Server-side extraction via pdfjs-dist with layout-aware line grouping and section heading detection.
2. Text → Claims — Llama models extract verifiable factual assertions per page. Claims are normalized, deduplicated (Jaccard + Dice trigram similarity), split into independent assertions, and merged when complementary.
3. Claims → Scored — Each claim is scored across 25+ materiality signals: named entities, metrics, temporal markers, attribution, financial terms, regulatory language, claim completeness, and table-relationship detection. Low-materiality fragments and derivative metrics are suppressed.
4. Scored → Evidence — A multi-step retrieval cascade runs per claim:
   • Deterministic queries (literal, metric-focused, entity, attribution, source-domain, official)
   • LLM-generated queries (when deterministic evidence is weak)
   • Wikipedia fallback
   • Source text fetching for top-credibility results
5. Evidence → Verdict — Retrieved sources are scored against a curated authority database (250+ domains across government, academic, news, and tech). A guardrail comparator pre-checks evidence before LLM synthesis. If the LLM call fails (rate limit, timeout), verdicts are derived from evidence guardrails alone.
6. Verdict → Report — Results stream to the client via SSE. Export as JSON (machine-readable) or PDF (formatted report with decision metadata).

Core Capabilities

Feature	Detail
Claim Extraction	Per-page extraction with multi-strategy JSON parsing, truncation repair, and Zod validation
Materiality Scoring	25+ feature signals — metrics, entities, temporal markers, attribution, section importance, claim completeness
Evidence Retrieval	Cascading search: Tavily → Serper → Mojeek (scrape) → Wikipedia API, with result caching and source text fetching
Source Authority	250+ curated domains across 6 tiers (government → academic → global news → official platforms → tech media → low-authority)
Verdict System	Verified · Inaccurate · False · Unverifiable — with confidence scores, corrected facts, and decision path transparency
Decision Metadata	Per-claim: decision_path, evidence_status, retrieval_status, reason_codes, comparator_verdict, search_query_count, duration_ms
Knowledge Fallback	When search engines return zero results, parametric model knowledge is used with a confidence cap of 65%
Export	JSON report (full metadata) or PDF report (formatted with @react-pdf/renderer)
Model Selection	Llama 4 Scout 17B (fast) or Llama 3.3 70B (thorough) via Groq's OpenAI-compatible API
Streaming	Server-Sent Events for real-time claim-by-claim progress with batch tracking

Stack

Layer	Technology
Framework	Next.js 15 (App Router, React 19)
Language	TypeScript 5
LLM Provider	Groq (OpenAI-compatible API)
Models	meta-llama/llama-4-scout-17b-16e-instruct, llama-3.3-70b-versatile
PDF Parsing	pdfjs-dist (legacy build, server-side)
PDF Export	@react-pdf/renderer
Schema Validation	Zod 4
Search APIs	Tavily, Serper, Mojeek (scrape), Wikipedia
Styling	Tailwind CSS 4, Radix UI primitives
Animations	Framer Motion
Testing	Vitest, Playwright
CI	GitHub Actions (lint → typecheck → test → build)
Deployment	Vercel (Node.js runtime, 60s function timeout)

Installation

git clone https://github.com/de2pressed/2see.git
cd 2see
npm ci

Environment Variables

Create a .env file from the example:

cp .env.example .env

Variable	Required	Description
GROQ_API_KEY	Yes	Groq API key for Llama model access. Get one free →
OPENAI_API_KEY	No	Optional fallback if GROQ_API_KEY is not set
TAVILY_API_KEY	Yes	Tavily search API for semantic evidence retrieval. Get one →
SERPER_API_KEY	No	Serper API for Google search results. Get one →

Minimum viable setup: Only GROQ_API_KEY is required. Without Tavily or Serper, evidence retrieval falls back to Mojeek web scraping and the Wikipedia API.

Usage

npm run dev

Open http://localhost:3000. Select a model, upload a PDF, and start analysis.

API Endpoints

Endpoint	Method	Description
/api/check-key	GET	Validates the configured Groq API key
/api/extract-claims	POST	Accepts a PDF (multipart/form-data), returns extracted + scored claims
/api/verify-claims	POST	Accepts claims JSON, returns SSE stream of verification results
/api/export-report	POST	Accepts a full report JSON, returns a formatted PDF

Report Evaluation

Compare a verification report against a golden-truth file:

node scripts/evaluate-report.mjs golden.json exported-report.json

Outputs verdict match rates, missing claims, forbidden verdict hits, source authority distribution, and average per-claim duration.

Verification Output

Each verified claim includes:

{
  "claim_id": "claim-1",
  "claim": "Global AI market reached $196.6 billion in 2023.",
  "type": "financial",
  "verdict": "Verified",          // Verified | Inaccurate | False | Unverifiable
  "confidence": 88,               // 0–100
  "explanation": "Multiple high-authority sources confirm...",
  "corrected_fact": "",            // populated when verdict is Inaccurate or False
  "sources": [
    {
      "title": "AI Market Size Report",
      "url": "https://...",
      "domain": "statista.com",
      "credibility": "High",       // High | Medium | Low
      "snippet": "...",
      "retrieved_at": "2025-01-15T..."
    }
  ],
  "decision_path": "llm",         // guardrail | llm | fallback | knowledge
  "evidence_status": "direct",    // direct | related | weak | absent | conflicting | technical_failure
  "retrieval_status": "searched", // not_needed | searched | fallback_searched | exhausted | quota_limited | technical_failure
  "reason_codes": [],
  "duration_ms": 4200
}

Project Structure

2see/
├── app/
│   ├── api/
│   │   ├── check-key/         # API key validation
│   │   ├── extract-claims/    # PDF upload → claim extraction
│   │   ├── verify-claims/     # Claim verification (SSE stream)
│   │   └── export-report/     # Report → PDF generation
│   ├── globals.css            # Design tokens (oklch palette, shimmer animations)
│   ├── layout.tsx             # Root layout, fonts, metadata
│   └── page.tsx               # Entry point → VerificationApp
├── components/
│   ├── verification-app.tsx   # Main UI (1400+ lines — upload, progress, results, filters, export)
│   ├── model-selector.tsx     # Llama model picker
│   └── ui/                    # Badge, Button, Progress (Radix-based)
├── services/
│   ├── openai.ts              # LLM calls, search cascade, evidence processing, verdict synthesis (~2800 lines)
│   ├── pdf.ts                 # PDF text extraction with layout-aware line grouping
│   └── verification.ts        # Batched verification orchestrator with retries
├── utils/
│   ├── claims.ts              # Materiality scoring, deduplication, claim merging (~800 lines)
│   ├── sources.ts             # Domain authority database, credibility scoring, evidence/hallucination risk
│   ├── ai.ts                  # JSON extraction, sanitization, truncation repair, Zod parsing
│   ├── chunk-text.ts          # Text chunking for long documents
│   ├── async.ts               # Concurrency-limited map, delay helper
│   └── files.ts               # PDF validation, filename sanitization
├── lib/
│   ├── schemas/index.ts       # Zod schemas for all data types
│   ├── models.ts              # Model definitions and selection helpers
│   ├── llm.ts                 # Shared completion parameters
│   ├── claim-cache.ts         # SHA-256 content-addressed claim cache
│   ├── search-fixtures.ts     # Test fixture loader for search results
│   └── utils.ts               # cn() (clsx + tailwind-merge)
├── types/
│   └── report.ts              # Re-exported schema types
├── hooks/
│   └── use-model-selection.ts # LocalStorage-backed model preference
├── scripts/
│   └── evaluate-report.mjs    # Golden-truth report evaluation harness
├── tests/                     # Vitest unit tests + fixtures
├── .github/workflows/ci.yml   # CI: lint → typecheck → test → build
└── .env.example

Limitations

• Free-tier rate limits. Groq's free tier has token-per-minute and request-per-minute caps. Batch sizes and delays are tuned for this, but large documents may hit throttling.
• Image-based PDFs. Text extraction requires selectable text. Scanned/image-only PDFs will fail with a clear error.

Roadmap

• Batch PDF processing
• Persistent report storage
• Additional LLM provider support
• Configurable claim extraction rules
• Source page screenshot capture
• Claim diff across document versions
• API-first mode (headless verification)

Development

npm run dev          # Start dev server
npm run lint         # ESLint
npm run typecheck    # TypeScript strict check
npm test             # Run Vitest suite
npm run build        # Production build

Contributing

Contributions welcome. Open an issue first for non-trivial changes.

1. Fork the repository
2. Create a feature branch (git checkout -b feat/my-change)
3. Commit with clear messages
4. Open a pull request against main

All PRs must pass the CI pipeline (lint, typecheck, test, build).

License

MIT