Generative AI application stack and providing long term memory to LLMs
summary
Claims that integrated vector+operational storage reduces ETL and improves retrieval accuracy, demonstrated by using MongoDB Atlas vector search to store embeddings alongside source data and metadata, plus hybrid retrieval (vector + BM25) with reciprocal-rank fusion for better results in jargon-heavy domains. Orchestration uses Microsoft Semantic Kernel wired to an embedding service (Azure OpenAI) and MongoDB memory store to build a RAG pipeline. Notes that search nodes running on Microsoft Cloud scale independently (2–32 nodes) and are still rolling out.
AI-written summary · about
chapters (7)
- 0:00Introduction to generative AI
- 2:51The modern application stack
- 5:01Vector databases in MongoDB
- 8:02Advanced querying capabilities
- 11:25Retrieval optimization
- 15:01Enterprise model requirements
- 17:44Orchestration with Semantic Kernel
related talks
26mBuilding generative AI–enriched applications with AWS & MongoDB Atlas
01h 16mAzure SQL ❤️ Open AI
12mBuilding Generative AI Apps with Azure Cosmos DB
50mHow to design and build AI applications with vector search using Azure OpenAI & Azure Cosmos DB
41mGenAI, vector/semantic/hybrid search, RRF, NLP, (...) and other buzzwords
55mSupercharged Search with Semantic Search and Vector Embeddings
44mVector Search Isn’t Enough
18m: Workforce Efficiency with AI-Driven Vector Search