Beyond the Vector DB
summary
Describes a production RAG that stays current by reembedding only changed chunks using checksum change detection and weekly refresh pipelines, tracks embedding model and chunk metadata with versioning, and applies semantic diffing via sentence transformers to catch subtle meaning changes. Uses cosine distance drift checks with an alert at cosine drift > 0.3 to decide rerank versus retrain, runs model lifecycle with MLflow, scores hits by combined relevance+recency, and exposes a dashboard plus human-in-the-loop alerts to avoid blind full reindexes.
AI-written summary · about
chapters (14)
- 0:00Introduction and speaker intro
- 2:13What are RAGs and their benefits
- 6:06Why freshness in RAG matters
- 8:13Traditional RAG architecture overview
- 9:19Step 1: Change detection with checksums
- 10:45Step 2: Embedding metadata and versioning
- 12:32Step 3: Semantic diffing for subtle changes
- 14:21Step 4: Indexing with intelligence
- 15:28Step 5: Embedding drift detection
- 17:36Step 6: Retraining workflow
- 19:42Step 7: Freshness aware retrieval
- 21:08Monitoring and metrics best practices
- 22:52Common mistakes to avoid
- 23:34Q&A session
related talks
33mVector Databases in the Age of AI: Embracing RAG for Enhanced Search Capabilities
06mWhat makes vector databases essential for RAG and AI apps?
15mMastering RAG and Vector Search with Elastic
36mUnleashing the Power of Vector Databases for AI-Driven Applications
15mEvaluating RAG Model Performance Metrics, Bias, and Interpretability
01h 35mTutorial: Build AI-Supercharged RAG Apps with a Vector Database
37mBuilding advanced RAGs with Elasticsearch OpenSearch
20mLayering every technique in RAG, one query at a time