Beyond RAG - Going From Search to Analytics on Unstructured Data with Aryn
summary
Builds an AI-powered unstructured data warehouse called Aryn that parses PDFs into images, tables, figures and text with a custom table extractor and VLMs, indexes chunks and metadata into OpenSearch, and runs semantic execution plans produced by Luna planner and executed with the Sycamore pipeline. Demo on ~100 NTSB reports retrieves 36 candidates and finds 2 bird strikes; internal bench shows 20/30 correct plans out of the box, 1.5–2.5x better segmentation and ~97% accuracy on a support task. Verifiability via execution lineage.
AI-written summary · about
chapters (12)
- 0:00Welcome to Aryn AI
- 1:04RAG limits and deep analytics
- 5:40Enterprise unstructured data use cases
- 8:48Agentic document intelligence
- 9:51Aryn’s platform architecture
- 10:35Live demo of analytics system
- 23:39Challenges in parsing documents
- 26:30Sycamore: Agentic dataflow for ETL
- 30:23Luna: Natural language query planning
- 32:29Verifiability and human-in-the-loop
- 34:15Conclusion and future directions
- 35:21Q&A on ambiguity handling
related talks
01h 28mFrom RAG to Riches Building Enterprise Search Systems with LLMs
32mRevolutionizing ETL for RAG with
23mNo RAGrets
29mRAG for Data Engineers: Why Your Metadata Is the Missing Piece
21mHow RAG is Redefining AI Capabilities
15mGraphRAG-Powered AI Agents: Transforming Corporate Metadata into Intelligent Knowledge
37mBuilding advanced RAGs with Elasticsearch OpenSearch
36mDon't tell me RAG is easy