Hybrid retrieval & agentic RAG — graphs/KG, failure postmortems, freshness patterns

LycheeMemory V2 (86% token reduction), RAG latency optimization via retrieval score margin routing, WisPaper citation verification, RAG loop dispatcher pattern. Production RAG pipeline for 10M+ docs. New today: NLP-to-MCP pipeline for EduTech RAG using LangGraph multi-agent orchestration; Multi-stage neural networks with learnable intermediate steps for production RAG under budget constraints; LLM/framework selection case study includes Graph-RAG, MCP, and formal constraints; RAG Pipeline Labs with security labs; RAG architecture full pipeline failures. Also: Production agents ditching pure RAG for context caching (68% latency reduction, 80% cost cut with hybrid routing). Today's reading adds: 6 Decisions That Make or Break Your RAG Pipeline; Detailed RAG cost breakdown ($1,308/month at 10M tokens/day, generation dominates 61%); Governed RAG pattern with graph-based control flow and human-in-the-loop; Loop engineering for resilient RAG pipelines; Retrieval Is a Pipeline: Sentence-Aware Chunking, Hybrid Search, Reranking; Making the Knowledge Layer a Graph You Actually Traverse; RAG AI Development for Enterprise Applications 2026 Guide; Azure Functions RAG pipeline patterns (fan-out/fan-in, APIM gateway). Today's reading adds: Graph architecture grounds RAG with structural constraints, reducing hallucinations; Multi-hop retrieval primer (DSPy, Baleen, Beam Retrieval, Self-RAG); FastEmbed vs SIE comparison for in-process vs shared inference; Production RAG beyond vector search (hybrid, graph, agentic query planning).