PDFs - When a thousand words are worth more than a picture (or table).
summary
Claims a vision-driven pipeline that uses multimodal models to turn PDF visuals into descriptive text chunks, improving retrieval vs raw parsing or reverse-engineering. Defines three chunking rules: complete scope, link complementary information, and accommodate varying complexity. Shows an experiment using OpenAI embeddings and cosine similarity where descriptive decomposition consistently yields higher overlap with organic user queries. Demonstrates converting tables, schematics and medical images into fluid, evidence-linked descriptions and recommends forcing evidence output to reduce hallucination and preserve source integrity.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:37RAG pipelines with PDFs
- 4:02PDFs for semantic extraction
- 6:55Chunking in RAG
- 9:21Multimodal models for PDFs
- 12:56Decomposing tables
- 16:49Table decomposition evaluation
- 21:25Rationale for images
- 25:42Key takeaways
- 26:24Q&A
related talks
21mStructuring the Unstructured
22mStructuring the Unstructured : Empowering the AI Workflows
26mAdvanced Document Parsing for AI Workflows
30mFrom Unstructured Docs to Graph Intelligence: A Framework for Building Graph+Embedding Pipelines
20mUltraviolet: Turn Hidden Document Data into an AI Advantage
32mRevolutionizing Automation: Retrieval-Augmentation with LLMs
22mContexts & Machines: How Document Parsing Shapes RAG results
01h 19mLearn to Unlock Document Intelligence with Open-Source AI