Aug 16, 2026 · ~2 min

Document AI OCR

Document AI for HWP and PDF notices · key-field F1 0.985 · about 2,000 notices/day

Document AI · Information Extraction

Built withPython · HWP/HWPX · PDF · Gemini 2.5 Flash-Lite · ThreadPool · Deterministic parsing

Document AI pipeline from direct download and magic-byte format routing through HWP and PDF parsers, source-field copying, deterministic numeric parsing, and recoverable batch controls — Browserless format routing with a strict boundary between language extraction and numeric calculation

WHY

Key inputs for model training and joint-bid matching, such as discipline shares and construction-capacity criteria, existed only inside attached HWP and PDF notices rather than in the listing API. Manually opening about 2,000 new notices each day could not support real-time feature creation or client matching. The required system had to structure those documents automatically while keeping every amount and ratio reproducible.

Key-field extractionF1 0.985

HWP/HWPX and PDF procurement notices

Daily throughputAbout 2,000

New notices processed automatically

Numeric boundaryDeterministic

Amounts, ratios, and units validated in code

HOW

Separate document recovery, field copying, and arithmetic

Each stage has one responsibility, making extraction errors observable and numeric outputs reproducible.

01Direct download
Fetch the source URL without browser automation and retain the original bytes.
02Format routing
Use magic bytes to distinguish HWP, HWPX, PDF, and ZIP rather than trusting the filename.
03Text recovery
Apply rhwp, hwp5, or liteparse through format-specific paths with bounded retries.
04Source-field copy
Ask Gemini 2.5 Flash-Lite to copy fields from the recovered text without calculation.
05Deterministic calculation
Parse amounts, ratios, and units in code, then append structured output and error logs.

RESULT

Measured extraction batch

Measure	Result	Why it matters
Documents	199 / 199	No failed extraction in the measured batch
Content	2,140,021 tokens	Large documents preserved instead of reduced to snippets
License-ratio GT	About 90%	Deterministic ratio handling after source-field extraction
Documented cost	About $0.82	Batch-scale use of a lightweight model

Document extraction result cards showing 199 of 199 documents extracted, 2.14 million tokens processed, about 90 percent license-ratio agreement, and documented batch cost of about 82 cents — Anonymized measured batch · HWP/HWPX, PDF, and ZIP documents

Contribution

Designed the direct-download, magic-byte routing, parser fallback, and deterministic numeric-processing boundary.
Built concurrent batching with stage-specific timeouts, retries, incremental CSV output, and structured error attribution.

Evidence

Private project