RAG and Voice Agent Applications
summary
Argues that production voice agents require a three‑layer memory: short LLM context, working memory via graph traversal, and long‑term graph storage; implements hybrid RAG combining vector similarity and Neo4j‑style multihop graph queries, stores ~2.5M edges to support ~250k candidates, uses confidence‑weighted edges with timestamps and a temporal decay halflife, reports up to ~87% context relevance versus vector‑only, and targets sub‑200ms responses for real‑time voice interactions.
AI-written summary · about
chapters (9)
- 0:00Introduction and background
- 1:27Limitations of vector search
- 3:35Three-layer memory architecture
- 5:44Hybrid RAG system design
- 8:56Hybrid RAG performance gains
- 9:36Confidence-weighted edges
- 11:21Temporal decay in memory
- 12:46Scaling voice agents
- 13:32Future vision and takeaways
related talks
27mWhen Vectors Aren’t Enough: Unlocking AI’s Hidden Connections with Graphs
29mThe GraphRAG System That Asks Back: Intelligent Probing with Contextual Answers
30mAgentic RAG with Knowledge Graphs Explained
15mRAG Beyond Chatbots
23mConnecting the Dots With Context Graphs
20mHybridRAG: A Fusion of Graph and Vector Retrieval
15mMemory-Augmented AI Agents: Enhancing Long-Term Reasoning and Context Awareness
28mEnhancing LLMs with Graph Technology