bRAG-langchain: five notebooks that climb from a basic RAG pipeline to ColBERT and Cohere

full_basic_rag.ipynb is the boilerplate entry point

The repository puts its shortcut at the root rather than burying it, full_basic_rag.ipynb, described as boilerplate starter code of a fully customizable RAG chatbot for anyone who wants to jump straight in. Around it sits the actual curriculum, five numbered notebooks under notebooks/ that move from introductory setup to multi-querying and custom RAG builds. The instruction to run files in a virtual environment is repeated before the notebook list, and the Getting Started section makes that concrete with per-platform Python install steps. The primary language registered on GitHub is Jupyter Notebook, not Python, which matches what the project is, a sequence of executable lessons rather than an application or a package. The companion site bragai.dev is tagged as launching soon.

Notebook one lays the baseline pipeline

The first notebook, 1_rag_setup_overview.ipynb, walks through environment setup with the necessary libraries and API setups, initial data loading with basic document loaders and preprocessing, embedding generation using various models including OpenAI's embeddings, a vector store configured with ChromaDB or Pinecone for similarity search, and finally a basic retrieval and generation pipeline that serves as the baseline everything later builds on. That closing phrase matters for how the curriculum reads, the first notebook deliberately produces the simplest thing that works, and each later notebook measures itself against it, notebook two even carries a comparison and analysis section that contrasts multi-query results with the single-query pipeline from this baseline.

Routing decides which source answers the question

Notebook three, rag_routing_and_query_construction.ipynb, is where the pipeline stops treating retrieval as one flat search. Logical routing implements function-based routing that classifies user queries to appropriate data sources based on programming languages. Semantic routing uses embeddings and cosine similarity to direct a question to either a math or a physics prompt. Query structuring defines a structured search schema for YouTube tutorial metadata, enabling filtering by view count or publication date, and structured search prompting has the LLM generate database queries from user input, linked into vector stores for retrieval. The notebook is a catalogue of dispatch strategies, each shown as working code against a concrete domain, source code questions, math versus physics, and YouTube metadata.

Indexing past one vector per document

Notebook four, rag_indexing_and_advanced_retrieval.ipynb, attacks the assumption that a document equals a single embedding vector. It opens with a preface pointing to external resources on document chunking, then sets up multi-representation indexing for documents carrying different embeddings and representations, stores document summaries in InMemoryByteStore alongside parent documents, and wires a MultiVectorRetriever over the result. Two named models get dedicated treatment, RAPTOR as an indexing and retrieval approach with links to in-depth resources, and ColBERT for token-level vector indexing that captures contextual meaning at a fine-grained level, demonstrated by retrieving information about Hayao Miyazaki from Wikipedia with the ColBERT retrieval model.

RAG-Fusion, reciprocal rank fusion and Cohere re-ranking

The final notebook, rag_retrieval_and_reranking.ipynb, assembles the system with a stated focus on scalability and optimization. It chunks documents for indexing, then uses RAG-Fusion, a prompt-based approach that generates multiple search queries from a single input question, and merges the resulting retrieval lists with Reciprocal Rank Fusion for improved relevance. A retriever and RAG chain built on the fused rankings pulls contextually relevant information to answer queries. Re-ranking gets two angles, a demonstration of Cohere's model for contextual compression and refinement, plus exploration of CRAG and Self-RAG retrieval approaches with linked examples, and a closing section of resources on how long-context retrieval affects RAG models.

Python 3.11.11 pinned, with a symlink repair step

The Getting Started section prefers Python 3.11.11 and scripts the whole path:

git clone https://github.com/bRAGAI/bRAG-langchain.git
cd bRAG-langchain

python3.11 -m venv venv

source venv/bin/activate

Windows uses venv\Scripts\activate instead. The interesting part is step three, an explicit repair for virtual environments that default to a newer interpreter, the section names Python 3.13 as the example. It verifies with python --version, falls back to invoking python3.11 directly, and offers a symbolic link, ln -sf $(which python3.11) $(dirname $(which python))/python, so the bare python command resolves to 3.11 inside the environment. Platform prerequisites are scripted too, brew install [email protected] on macOS, sudo apt install python3.11 python3.11-venv on Linux, and the installer from Python.org with its Add Python to PATH checkbox on Windows.

One .env.example, four vendors, twenty-one requirements

The repository's dependency file lists twenty-one pinned or bare requirements, the six langchain-prefixed packages, langchain, langchain-community, langchain-core, langchain-openai, langchain-pinecone and langchain-cohere, plus pydantic>=2.7.0, pinecone-client, chromadb, cohere, tiktoken, numpy, beautifulsoup4, pypdf, ragatouille, python-dotenv, ipykernel, langchainhub, and a YouTube toolchain of youtube-transcript-api under 1.0.0, pytube and yt_dlp. The .env.example at the root names every account you must hold, an OpenAI key from platform.openai.com, LangSmith tracing with LANGCHAIN_TRACING_V2=true, an endpoint, key and project name, Pinecone with PINECONE_INDEX_NAME, PINECONE_API_HOST and PINECONE_API_KEY, and a Cohere key. The bill for the curriculum is therefore four vendor relationships before the first notebook runs.

Editorial conclusion

Work through bRAG-langchain if you learn best by executing a graded sequence of notebooks and are prepared to hold OpenAI, Pinecone and Cohere API keys, since the .env.example wires all three vendors plus LangSmith tracing. Skip it if you want an installable library rather than a course, or a JavaScript stack, because everything here is Python pinned to 3.11.11. Verify the license before commercial reuse, GitHub reports NOASSERTION even though a LICENSE file sits at the repository root, and confirm the bragai.dev site has launched if you prefer prose documentation over notebooks.

Frequently asked questions

Official sources

1. bragai/bRAG-langchain on GitHub
2. Issues
3. Project website
4. README