Awesome AI Agents › Courses and Tutorials

GiovanniPasq/agentic-rag-for-dummies

⭐ 4368 Jupyter Notebook added to this list on 2026-03-01 repository created 2025-10-13

agentic-rag-for-dummies is an MIT-licensed teaching repository and reference implementation showing how to build an agentic retrieval-augmented generation system on LangGraph with minimal code. It offers two entry points: an interactive Jupyter notebook, also runnable in Google Colab, that walks through the concepts step by step, and a modular project in which the LLM provider, embedding model, PDF converter and agent workflow can each be adapted independently. Documents are indexed hierarchically, with parent chunks as bounded large sections split on Markdown headers and child chunks as small fixed-size pieces derived from them, combining the search precision of small chunks with the contextual richness of large ones. Query handling runs in four stages. First, a rolling summary plus recent conversation history preserves continuity without letting context grow indefinitely. Second, a clarification stage resolves references, splits multi-part questions into focused sub-queries, detects unclear input, rewrites queries for retrieval and pauses for human input when clarification is needed. Third, retrieval runs as a multi-agent map-reduce: LangGraph spawns one parallel agent subgraph per sub-query, each searching child chunks, fetching parent chunks for context, self-correcting when results are insufficient, compressing context to avoid redundant fetches and falling back gracefully when the search budget is exhausted. Fourth, an aggregation stage merges the agent responses into a single coherent answer. Qdrant serves as the vector store. The runnable application is Ollama-first and the README recommends models of eight billion parameters or larger for reliable tool calling, but chat model initialisation can target any LangChain-supported provider, with examples given for OpenAI, Anthropic and Google. Langfuse provides observability over LLM calls, tool usage and graph execution, and RAGAS metrics evaluate retrieval and answer quality. It requires Python 3.11 or newer and suits developers learning agentic RAG patterns or bootstrapping their own pipeline.

https://github.com/GiovanniPasq/agentic-rag-for-dummies

raglanggraphtutorialretrievalqdrant