Contexts & Machines: How Document Parsing Shapes RAG results
summary
Experiment parses five ~100‑page technical reports, generates ~200 LLM Q&A pairs with RAAS, and compares pypdf, PDFact, unstructured and llama pars combined with rule‑based and semantic chunking (chunky). Key findings: rule‑based chunking outperforms semantic out of the box, larger chunks improve multi‑chunk (multi‑hop specific) answers while shorter chunks help single‑hop queries, and no parser wins overall. Document type guides parser choice (unstructured for multi‑column, PDFact for dense text, pypdf for form‑like pages).
AI-written summary · about
chapters (10)
- 0:00Introduction to RAG and document parsing
- 2:06Challenges of PDF parsing
- 6:00Experiment setup and data generation
- 8:46Chunk size and parsing results
- 11:21Parser selection by document type
- 14:09Key takeaways and recommendations
- 15:39Q&A: Handling headers and hierarchy
- 18:11Q&A: Evaluation of chunking methods
- 20:21Q&A: Multi-chunk question generation
- 21:28Closing and acknowledgments
related talks
26mAdvanced Document Parsing for AI Workflows
18mLLMs and GenAI: A Guide to Building Your Own RAG Model for Document Interaction
21mBeyond RAG: Building Robust Knowledge Frameworks for Complex Multi-document Retrieval for Enterprise
32mBuilding Multi-Agent RAG Systems
01h 34mHow to Construct and Assess Your First LLM Based RAG System
32mBuilding an Advanced Knowledge Assistant
21mStructuring the Unstructured
22mStructuring the Unstructured : Empowering the AI Workflows