ganoro/df1ef9f0-3138-4b76-8be9-a0e40bc4ccef/claude-plugin/skills/data-aggregation/document-parser

Document Parser

Extract structured data from documents in the workspace. This skill is called during the AGGREGATE phase of the ACSR lifecycle.

Supported Formats

| Format | Method | Key Libraries | |--------|--------|--------------| | PDF | Python: pdfplumber or PyMuPDF | Tables, text blocks, page-level extraction | | XLSX | Python: openpyxl | Named ranges, multiple sheets, formulas | | DOCX | Python: python-docx | Paragraphs, tables, headers | | CSV | Python: pandas | Tabular data |

Extraction Workflow

Step 1 — Discover Files

find /workspace -type f \( -name "*.pdf" -o -name "*.xlsx" -o -name "*.docx" -o -name "*.csv" \) | head -100

List all parseable files. Present the inventory to guide prioritization.

Step 2 — Classify Each Document

Before parsing, classify the document type:

• Financial Statements — P&L, balance sheet, cash flow
• Fund Reports — quarterly/annual LP reports, capital account statements
• Valuation Reports — third-party or internal valuations
• Legal Documents — LPA, side letters, subscription agreements
• Data Room Materials — due diligence documents, CIM, management presentations
• Market Data — industry reports, benchmarks, comparable transactions
• Correspondence — emails, memos, meeting notes

Step 3 — Extract Data

Write a Python extraction script tailored to the document type. Example for PDF:

import pdfplumber
import json

def extract_pdf(filepath):
    results = []
    with pdfplumber.open(filepath) as pdf:
        for i, page in enumerate(pdf.pages):
            # Extract tables
            tables = page.extract_tables()
            for table in tables:
                results.append({
                    "type": "table",
                    "page": i + 1,
                    "data": table
                })
            # Extract text blocks
            text = page.extract_text()
            if text:
                results.append({
                    "type": "text",
                    "page": i + 1,
                    "content": text
                })
    return results

Step 4 — Structure Output

For each extracted fact, create a provenance record:

## Extracted: [description]
- **Value**: [value]
- **Source**: [filename], page [N] / sheet [name] / section [heading]
- **As-of Date**: [date if identifiable]
- **Confidence**: VERIFIED if from audited source, REPORTED otherwise
- **Raw Text**: "[exact text from document]"

Step 5 — Write Source Files

Save extracted data to _research/sources/source-NNN.md with:

• Document metadata (filename, type, date, pages/sheets)
• All extracted facts with provenance
• Extraction quality notes (any parsing issues, low-confidence extractions)

Handling Common PE/VC Documents

Capital Account Statements

Look for: beginning balance, contributions, distributions, management fees, carried interest, ending balance, IRR, TVPI, DPI, RVPI.

Quarterly LP Reports

Look for: NAV, portfolio company updates, cash flow summary, performance metrics, market commentary.

Financial Statements

Look for: revenue, EBITDA, net income, total assets, total debt, cash, working capital. Note the reporting period and currency.

Valuation Reports

Look for: methodology (DCF, comparables, precedent transactions), key assumptions (discount rate, exit multiple, growth rate), fair value conclusion, ASC 820 level classification.

Error Handling

• If a PDF is scanned/image-based, note it as a gap: "Document [X] requires OCR — not extracted"
• If tables are malformed, extract what's readable and flag quality issues
• If a file is password-protected, note it as a gap
• Never guess values from partially readable text — mark as gap instead