Cutting RAG Token Bloat: Compression Before Retrieval in Production Pipelines

Battle-tested patterns for compressing vector search results and tool outputs before LLM ingestion, with pgvector indexing and hybrid rank fusion code.