RAG Practice with LangChain & LangGraph

This repository contains my personal experiments and practice scripts for building Retrieval-Augmented Generation (RAG) systems using LangChain and LangGraph.

I use this space to test different vector stores, embeddings, and LLM integrations to understand various RAG architectures.

Current Projects

1. RAG-pinecone.py

A RAG implementation leveraging Pinecone as the serverless vector store.

• Model: DeepSeek (deepseek-chat)
• Embeddings: Ollama (mxbai-embed-large)
• Vector Store: Pinecone (Serverless)
• Features:
  • Loads text files from data_files/ directory.
  • Splits text into chunks.
  • Indexes chunks into a Pinecone index named fictional-universe.
  • Defines a custom tool search_retrieve for the agent.
  • Example query: "What is a shared universe?"

2. RAG-token-efficient.py

A local RAG setup using ChromaDB focused on efficient token usage.

• Model: Google Gemini (google_genai:gemini-2.5-flash)
• Embeddings: Ollama (mxbai-embed-large)
• Vector Store: ChromaDB (Persisted locally in vectors/)
• Features:
  • Loads a specific file (Marvel_Cinematic_Universe.txt).
  • Uses a standard LCEL chain (RunnablePassthrough, StrOutputParser) instead of an agent.
  • Simple prompt template for context-based answering.

Setup & Usage

1. Environment Variables: Create a .env file in the root directory with the following keys:

   PINECONE_API_KEY=your_pinecone_key
   DEEPSEEK_API_KEY=your_deepseek_key
   run_api_key=your_langsmith_key # Optional
   GOOGLE_API_KEY=your_google_ai_key

2. Ollama: Ensure you have Ollama running locally with the mxbai-embed-large model pulled:

   ollama pull mxbai-embed-large

3. Dependencies: Install the required packages (see pyproject.toml or install manually):

   pip install langchain langchain-openai langchain-community langchain-chroma langchain-pinecone pinecone-client python-dotenv langchain-deepseek

Future Plans

I plan to add more variations of RAG pipelines, including:

• GraphRAG implementations using LangGraph.
• Hybrid search (Keyword + Vector).
• Multi-modal RAG.
• Advanced retrieval strategies (Self-querying, Parent Document Retriever).