Llm

What separates a toy RAG from a production-grade one. Three pillars covered: chunking strategies (character vs token-based vs HTMLHeaderTextSplitter), advanced retrieval with hybrid search (BM25 + dense) and Reciprocal Rank Fusion, cross-encoder reranking with Flashrank, and automated evaluation using LLM-as-judge with 4 RAGAS-equivalent metrics.

Can we use historical user feedback to decide if a new RAG config is better before deploying it? A complete offline evaluation pipeline: labeled pair extraction, query expansion, LLM-as-Judge comparison, and metric computation with good_maintenance_rate and bad_change_rate.

A complete hands-on guide to LangGraph's core building blocks: StateGraph and custom graphs, state reducers with add_messages, checkpointing with InMemorySaver, human-in-the-loop with interrupt and Command, and dependency injection in tools with ToolRuntime (state, context, store).

How to extend LLM capabilities with Tools and Agents. Three progressive levels: defining tools with @tool, understanding the manual ReAct loop with bind_tools and ToolMessage, and automating everything with create_react_agent from LangGraph. Covers sequential vs parallel tool calls, stateless LLM behavior, and how agents manage conversation history automatically.

RAG solves two fundamental LLM limitations: hallucination and knowledge cutoff. A complete hands-on guide covering the full pipeline — document loading, text splitting, HuggingFace embeddings, InMemoryVectorStore, retriever, and a full LCEL RAG chain. Free, no API cost needed.