Document Classification

Document Classification Skill

This skill takes a PDF from input_data/, classifies it, abstracts metadata, and uses an orchestration script to validate the company ticker and produce a standardized markdown file in processing_data/.

Prerequisites

• Python 3.10+
• PDFs placed in input_data/ directory at the project root
• If a static file server is not running on localhost:8181 then ask the user to run .\tools\start_file_server.bat

DO NOT EVER start servers without human user.

Inputs

• One or more PDF files in input_data/

Outputs

• Renamed PDF in processing_data/: TICKER_DOCTYPE_YYYYMMDD_temp.pdf
• Markdown file in processing_data/: TICKER_DOCTYPE_YYYYMMDD_temp.md

---

Step-by-Step Instructions

Batch Mode: Repeat Steps 1–6 for EVERY .pdf file in input_data/. Track which files have been processed and continue until all are done. Report a summary at the end.

Step 1: Pick the Next PDF to Process

1. List all .pdf files in input_data/
2. If no PDFs exist (or all have already been processed), stop and inform the user: "No PDFs found in input_data/" or "All PDFs processed."
3. Pick the next unprocessed PDF in input_data/. Keep track of its filename.
4. Continue to Step 2 for this PDF.

Step 2: Read the PDF

1. Ensure a background process running the file server is active (.\tools\start_file_server.bat).
2. Open the PDF in the browser using the browser_subagent tool:
   • Navigate to the PDF via the local server, e.g., http://localhost:8181/input_data/{filename}
   • The browser renders PDFs natively with full visual fidelity
3. Read the first 1–3 pages to extract classification metadata (company name, ticker, document type, dates)
4. For large documents (10-K, 10-Q, analyst reports), you do NOT need to read every page — focus on the cover page and table of contents
5. Do NOT use PyPDF2 or other text extraction libraries — the browser approach preserves table layouts and handles scanned/image-based documents
6. If the PDF cannot be opened in the browser, stop and inform the user