Knowledge Share
Technical articles, tutorials, and insights
RAG Series (20): Enterprise RAG Architecture
Part 20 of the RAG series. Demo-grade RAG gives every user access to the same vector store. Enterprise deployments need three things — multi-tenancy (different companies/departments can't see each other's documents), access control (engineers can't retrieve HR docs, HR can't retrieve financial data), and a service layer (caching for repeated questions, rate limiting to prevent abuse). This article implements all three using Qdrant Collections for isolation, metadata filters for permissions, and a sliding-window rate limiter. Five scenarios verify everything works.
RAG Series (21): Performance Optimization — Faster and Cheaper
Part 21 of the RAG series. Every RAG request makes two API calls — Embedding and LLM — both metered by token and both adding latency. Four optimizations measured independently with real data: LLM response cache (5057ms → 0.8ms, 6068× speedup), Embedding cache (repeat texts at zero API calls), Semantic Cache (reuse answers for similar questions, but threshold calibration is non-trivial), async batch Embedding (2.87× speedup for 12 texts). Data speaks for itself.