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

Step 3: Classify the Document

From the PDF content read in Step 2, extract the following fields. Refer to skills/document_classification/resources/document_types.json for validation and exact naming of Document Types.

Fields to extract:

Anti-hallucination rules:

• ONLY extract information EXPLICITLY shown in the document text
• DO NOT invent, infer, or assume company names, tickers, or dates
• If information is not visible, use null
• DO NOT use external knowledge to fill in missing information at this step

Step 4: Extract Dates

Extract each date field carefully, paying close attention to the document context.

Fiscal Year Awareness: Before assigning time_period, check skills/document_classification/resources/fiscal_year_map.json to see if this company has a non-standard fiscal year. Many retail and tech companies (e.g., LULU, WMT, COST, NKE, BABA) have fiscal years ending in months other than December. This shifts which calendar months map to which fiscal quarters.

• If the ticker IS in the map, use the fiscal_year_ends entry to determine the correct fiscal quarter from the period_end_date.
• If the ticker is NOT in the map, assume a standard calendar fiscal year (ending December 31).

4a: Extract document_date

• The date the document was published or released
• Format: YYYY-MM-DD

4b: Extract time_period

• The fiscal reporting period
• Format: Q1 YYYY, Q2 YYYY, Q3 YYYY, Q4 YYYY, or FY YYYY
• Important: Use the company's fiscal calendar, not the calendar year. For example, LULU's fiscal year ends in January, so a period ending Jan 28, 2024 is Q4 FY2023, not Q1 2024.

4c: Extract period_end_date

• The date the financial period ended (quarter end or fiscal year end)
• Format: YYYY-MM-DD

Step 5: Execute Orchestration Script

Run the python orchestration script to validate and finalize the classification. This script will automatically validate the ticker against Yahoo Finance, rename strings according to document type abbreviations, move the file, and create the final markdown file.

Execute the following terminal command with the correctly formatted strings from Steps 3 and 4:

bashCopy code
uv run python skills\document_classification\scripts\process_classification.py --filename "FILENAME.pdf" --company_name "Extracted Company Name" --ticker "TICKER" --document_type "document_type" --document_date "YYYY-MM-DD" --time_period "time_period" --period_end_date "YYYY-MM-DD" --confidence "high"

If it succeeds, output the success to the user. If there are more PDFs remaining in input_data/, return to Step 1 for the next file. Otherwise, print a final summary of all processed documents and finish the run.

💡 First-File Scenario: If the document is the first one being processed for a ticker (no output_data/TICKER exists), Phase 5 (Qualitative Assessment) will require you to manually initialize the ticker directory and TICKER_metadata.md file.

Step 6: Reflection — Use LLM Knowledge (Fallback)

If the script in Step 5 throws an error (e.g. invalid ticker):

1. Ask the LLM: "Given this company name and document context, what is the correct stock ticker?"
2. The LLM IS allowed to use its knowledge here (unlike Step 3)
3. If a new ticker is found, re-run Step 5 with the new ticker.
4. If it still fails, ask the human user for the correct ticker.

Error Handling

• If PDF cannot be opened in the browser → Inform user, skip this PDF
• If classification fails → Retry once, then inform user
• If ticker validation completely fails → Ask human user for the correct ticker
• If date extraction fails → Set failed fields to null, proceed with available data

Changelog

• 2026-05-06: Updated example command to use uv run per project standards. Added note about first-file initialization for qualitative documents.