Notes
Short technical write-ups from my projects.
Hybrid retrieval with BM25 + dense + cross-encoder reranking in my Agentic RAG
How search_documents combines keyword and vector search, why it fuses by rank, and where the reranker sits.
Refusing instead of guessing: the grounding gate in my multimodal chatbot
How a three-zone dense gate, grader reason codes and a web-result quality check keep answers tied to sources.
How I cut LLM token cost by 70% at Mapin
A high-level look at the two levers behind the reduction: custom text chunking and model routing on OpenRouter.