Documents Meet LLMs: Tales from the Trenches
summary
Pipeline extracts marketing claims from brochures, uses OCR (Amazon textract, Azure AI document intelligence or mark it down) and prompts for XML output or JSON with enforced schema. Grammar‑constrained decoding and cloud sonet 4.5 reduce errors, retries use tenacity and instructor and fallbacks switch providers. Postprocess escapes characters, apply regex checks, and measure reformulations with Levvenstein distance. Logging plus metrics surface misformats, missing vs hallucinated claims, and whether the structured extractor actually works in production.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 2:12Processing text documents
- 13:39Structured output challenges
- 21:05Evaluation pitfalls
- 24:15Q&A
related talks
10mSteering Semantic Data Processing With DocWrangler
21mStructuring the Unstructured
36mTechniques to avoid common LLM pitfalls
28mSupercharge Information Extraction with Vision-Based LLMs
01h 00mComplex Data Ingestion with Open Source AI
57mLLMs gone wild
26mAgentic Fundamentals: How language models and code can communicate
49mNext-Gen Documentation with AI