RFQ extraction exposes local LLM limits

The smart move here is to stop treating the LLM as the whole parser and use it only after deterministic extraction, chunking, schema validation, and retry loops have done most of the work.

• –RFQ extraction usually needs layout-aware parsing for tables, not just Markdown conversion plus a prompt
• –Smaller local models can miss fields, hallucinate structure, or degrade on long context, especially on consumer GPUs
• –A production pipeline should combine OCR/table extraction, field normalization, constrained JSON output, validation, and human review for low-confidence cases
• –The privacy case for local inference is real, but the economics may favor cloud or larger hosted models if throughput and accuracy matter