Fusing NLP and Graph: Building a Conversational Agent for Enriched Financial Data
summary
Built an end-to-end pipeline that turns 400‑page PDFs into a validated knowledge graph by combining LLM extraction, graph matching, and a reasoning agent. Chunking by chapter raises extractions from ~15 to ~150 (10×) and deduplication collapses those to ~80 distinct entities; a golden baseline refreshes every 8 hours. The stack uses Gemini 3.5 Pro for extraction, Neo4j for deduplication, graph edit distance, Jaccard overlaps and vector retrieval, plus an agent layer with human‑in‑the‑loop validation and document‑level provenance.
AI-written summary · about
chapters (12)
- 0:00Introduction and motivation
- 1:13Business context and knowledge graph
- 4:02Data pipeline overview
- 6:34NLP enrichment with chunking
- 9:47Prompt optimization method
- 11:53Entity resolution and validation
- 15:06Conversational agent introduction
- 17:24Human-in-the-loop validation
- 20:03Graph RAG and agent reasoning
- 22:52Unified reasoning platform
- 25:48Conclusion and future outlook
- 26:29Q&A session
related talks
28mAgentic GraphRAG: Multi-Agent Knowledge Graph
29mAgentic GraphRAG: Multi-Agent Knowledge Graph Construction for Research Teams
30mFrom Unstructured Docs to Graph Intelligence: A Framework for Building Graph+Embedding Pipelines
15mGraphRAG-Powered AI Agents: Transforming Corporate Metadata into Intelligent Knowledge
28mBuilding Knowledge Graph-Based Agents with Structured Text Generation
30mBuilding Knowledge Graphs With LLMs
32mBuilding an Advanced Knowledge Assistant
30mAgentic AI for Enterprise Knowledge Graph Management