Summary
Fahd Mirza tests Lift, a 9-billion parameter model purpose-built for schema-constrained PDF extraction, deployed locally on Ubuntu with a single NVIDIA RTX 6000 GPU (48 GB VRAM). Unlike general-purpose LLMs that produce unpredictable or malformed JSON, Lift uses schema-constrained decoding so outputs always conform to a provided JSON schema — missing fields return null rather than hallucinated values, and multi-page documents are processed in a single pass via rendered page images.
The video covers the full setup pipeline using vLLM (consuming approximately 41 GB VRAM at load), then runs two extraction tests. First, a messy AI-generated invoice with multiple addresses, a line-item table containing blank filler rows, and intentionally empty fields like PO number — Lift returns clean, correctly-typed JSON with no invented data and exactly three line items extracted. Second, a three-page multilingual ledger written in Chinese, Hindi, Arabic, Russian, Japanese, and several other languages, with field labels in each native language, mixed numeral systems, right-to-left text, and deliberate missing fields.
Mirza benchmarks the 9B model against hosted APIs and larger general models, finding it competitive for this specific task class at a fraction of the cost and on a single consumer-grade GPU. He notes the main practical friction: JSON schemas must be defined upfront, which may require a pre-processing pipeline for production use. The tool is served via the Lift vLLM helper command and the LLaVA PDF library.
📺 Source: Fahd Mirza · Published July 10, 2026
🏷️ Format: Hands On Build
