Simple RAG Pipeline with LangChain

A compact RAG implementation that reads .md and .txt files, splits them into chunks, generates embeddings via OpenAI, stores them in ChromaDB, retrieves relevant context for a user question, and sends the grounded prompt to a chat model. Built to demonstrate the full RAG loop without unnecessary abstraction.

Most RAG tutorials skip the practical details: chunking strategy, retrieval tuning, and how to inspect what the LLM actually received as context.

Build the simplest possible RAG pipeline that still shows the complete flow : ingestion → embedding → retrieval → generation : with observable inputs and outputs at each stage.

• File loader for local .md and .txt documents.
• Recursive text splitting with configurable chunk size and overlap.
• OpenAI embedding generation and ChromaDB storage.
• Similarity search retrieval with top-k results.
• LLM call with grounded context and source citations.

• Full pipeline from document folder to answer.
• Configurable chunk size and retrieval count.
• Printed context window for debugging.
• Source document citation in answers.

• OpenAI text-embedding-ada-002 for chunk embeddings.
• Cosine similarity search for retrieval.
• Chat completion with system prompt for grounded answering.

1. 1Documents placed in the input folder.
2. 2Script loads, chunks, and embeds all documents.
3. 3User asks a question via CLI.
4. 4System retrieves top-k relevant chunks.
5. 5Context + question sent to LLM.
6. 6Answer returned with source references.

• Choosing chunk size that balances detail with retrieval relevance.
• Handling documents that span multiple topics.
• Measuring retrieval quality without a labeled test set.

• Keep the pipeline simple and readable over optimizing retrieval metrics.
• Use OpenAI embeddings for reliability during development.
• Print the full context window for debugging instead of hiding it.

Working CLI-based RAG pipeline. Drop documents in a folder, ask questions, get grounded answers with source references.

• Add support for PDF ingestion.
• Add hybrid search (keyword + vector).
• Add evaluation script for retrieval quality.