Financial RAG Pipeline

End-to-end structured financial data extraction from multi-page PDF reports.

IBM Docling → ChromaDB → CohereRerank → Gemini 1.5 Pro → Validated JSON

Architecture

Phase	Component	Purpose
1	IBM Docling	Parse PDFs → Markdown (text, tables, figures)
2	all-MiniLM-L6-v2 + ChromaDB	Semantic chunking & vector storage
3	Cosine Top-500 → CohereRerank v3	Two-stage retrieval → Top 40–50 chunks
4	Gemini 1.5 Pro (LangChain)	Structured extraction with Pydantic validation
5	Validated JSON	50+ financial fields per company/period

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Configure Environment

cp .env.example .env
# Edit .env with your API keys

Required API keys:

• Google AI — for Gemini 1.5 Pro (Google AI Studio)
• Cohere — for CohereRerank v3 (Cohere Dashboard)

3. Run the Pipeline

# Basic usage
python main.py path/to/annual_report.pdf

# Custom output path
python main.py path/to/annual_report.pdf --output results.json

# Custom ChromaDB directory
python main.py path/to/annual_report.pdf --chroma-dir ./my_chroma_db

Output Format

The pipeline extracts 50+ financial fields per company/period into a validated JSON structure:

[
  {
    "company_name": "Example Corp Ltd",
    "period": "FY2024",
    "currency_unit": "INR Crores",
    "income_statement": {
      "revenue": 15234.5,
      "ebitda": 3200.0,
      "pat": 1850.0,
      "eps_basic": 45.2,
      ...
    },
    "balance_sheet": {
      "fixed_assets": 8500.0,
      "borrowings": 2100.0,
      ...
    },
    "cash_flow": {
      "cash_flow_from_operations": 2800.0,
      "cash_at_end": 1200.0,
      ...
    }
  }
]

Project Structure

Financial Document Analyzer/
├── main.py                 # CLI entry point
├── requirements.txt        # Python dependencies
├── .env.example            # Environment variable template
├── README.md               # This file
└── src/
    ├── __init__.py
    ├── models.py           # Pydantic schemas (50+ fields)
    ├── parser.py           # IBM Docling parsing
    ├── embeddings.py       # Chunking & ChromaDB storage
    ├── retriever.py        # Two-stage retrieval + CohereRerank
    ├── extractor.py        # Gemini 1.5 Pro extraction chain
    └── pipeline.py         # Full orchestration + error handling

Cost Estimation (per 500-page document)

Service	Usage	Approx. Cost
IBM Docling	500 pages (Local CPU)	$0.00
all-MiniLM-L6-v2	~500K tokens (Local CPU)	$0.00
CohereRerank v3	500 docs reranked	~$0.001
Gemini 1.5 Pro	~100K tokens	~$0.35
Total	—	~$0.35

Known Limitations

Limitation	Mitigation
Docling misses scanned text	Pre-process with external OCR
Chunk boundary splits table row	Increase overlap to 200 tokens for table-heavy docs
Gemini hallucination on sparse context	Pydantic validation + null count thresholds
Multiple companies in one chunk	Post-process: validate company_name per row
Units vary across documents	Enforce currency_unit field; normalise downstream