Skip to main content

Agentic RAG & Financial Data Extraction

To process corporate financial statements (IFS/BCTC) that carry strict mathematical constraints, PDFHUB deploys Agentic RAG. The system acts as a virtual financial analyst, cross-checking numbers and exporting clean datasets ready for empirical analysis.

1. Principles of Financial Agentic RAG​

Unlike standard RAG systems that split PDFs and search solely by vector similarity, PDFHUB's Agentic RAG follows a goal-oriented reasoning workflow:

User Query ──► Clarify Intent (Interpret financial metrics)
                  └──► Execute Plan (Query SQL/Cypher database or run Vector Search)
                         └──► Cross-Check (Verify accounting consistency)
                                └──► Synthesize (Return answers with citations)
  1. Intent Clarification: Recognizes target financial indicators (e.g., Debt-to-equity ratio, ROE, Cash flow from operating activities).
  2. Query Planning: The AI agent determines whether to fetch structured data using SQL/Cypher on previously parsed tables or run a vector search on the Notes to the Financial Statements.
  3. Evidence Collection: Locates the original tables and notes, preserving page coordinates and table IDs to establish a visual audit trail.
  4. Synthesis & Validation: Compares figures across the balance sheet, income statement, and notes to ensure mathematical consistency.

2. Accounting Logic & Mathematical Cross-Checking​

Financial statements require absolute precision. PDFHUB's RAG agent integrates automated cross-checking algorithms:

  • Aggregational Constraints: Verifies if the sum of sub-items matches the parent totals.
    • Total Assets = Current Assets + Non-Current Assets
    • Total Equity & Liabilities = Total Liabilities + Owner's Equity
  • Cash Flow Consistency: Matches the ending cash balance on the Cash Flow Statement with the corresponding cash indicators on the Balance Sheet.
  • Anomaly Logging: If discrepancies are found (due to rounding errors or parser issues), the system will:
    • Highlight the suspected cell values in red on the UI.
    • Log a warning outlining the mathematical mismatch.
    • Trigger the parser to rerun on those pages at a higher resolution.

3. Data Quality Reports & Export Formats​

Upon completion of extraction and validation, PDFHUB provides:

Data Quality Report​

  • Fill Rate: The percentage of schema fields populated.
  • Discrepancy Log: A list of mathematical mismatches along with their source page locations.

Data Export Options​

The system exports normalized financial data in formats suitable for quantitative analysis:

  • Parquet / DuckDB: High-performance formats that preserve column data types, ready to be loaded into Pandas (Python) or Plm (R).
  • CSV / XLSX: Standard sheet formats for quick imports into Stata, SPSS, or Microsoft Excel to run regressions.
  • Markdown / LaTeX: Clean Markdown tables to insert directly into academic research papers.