Visual Explainer: Document Classifier

Stage 1: Strip to Markdown & Take the Cover Slice

Why read 50 pages of holding tables when 98% of the document's legal identity is on page 1?

📑

1. Raw 50-Page PDF

Heavy, binary layout with images, fonts & tables (~150,000 chars)

📝

2. Stripped to Markdown

Pure clean plain text. Eliminates visual styling overhead

✂️

3. First 3,000 Chars Slice

Page 1 cover: contains Title, Date, ISIN, and statutory mandate

Interactive: Test Slice Length on Accuracy & Speed 3,000 characters (Optimal)

500 chars 1,500 chars 3,000 chars (★) Full Document (50k+)

Document Anatomy First 3k chars vs Rest of PDF

Cover (Kept)

49 Pages of Holdings & Tables (Discarded)

✓ Contains 98% of statutory titles & dates ✗ 50 pages of noisy accounting tables removed

Stage 2: The Rarity Filter (TF-IDF)

Word Importance = Frequency in Document × Rarity Across Library.

1. Frequency in This Document (TF): 18 times

2. Rarity Across 3,500 Library Files (IDF): Found in 100% of files

Final Mathematical Weight: 0.02 / 1.00

Stage 3: The Multi-Class Balance Scale (Logistic Regression)

Click clues to drop weights onto the scale. Watch how evidence physically tilts the decision.

Monthly: 50% Quarterly: 50%

Balanced: Both document types share equal likelihood.

Click words found in document header to add/remove evidence:

💡

Why "June 30" doesn't tilt the scale: Funds publish both June Monthly Factsheets and Q2 Quarterly Factsheets on June 30. The scale only tilts when unambiguous structural markers (like "quarterly report" or "1M return") are present.

Stage 4: Live Document X-Ray

Pick a real cover slice to see the detected features and points.

Top Detected Clues:

Result: Quarterly Factsheet

Confidence: 99.8%