Local RAG Playground — Chat With Your Documents In-Browser

This playground implements a complete Retrieval-Augmented Generation (RAG) pipeline entirely within your browser, ensuring maximum privacy. When you upload documents, the process begins: first, the system chunks your files into manageable segments. Next, each chunk is converted into numerical vectors using the all-MiniLM-L6-v2 embedding model. These embeddings capture the semantic meaning of your text.

When you ask a question, we perform a cosine similarity search against these stored vectors, retrieving the top 5 most relevant chunks (top-5 retrieval). Finally, the retrieved context is passed to an optional local Large Language Model (LLM) to generate a coherent answer. Crucially, this entire workflow—chunking, embedding, and querying—happens locally, meaning your documents never leave your device.

To maximize the accuracy of your RAG pipeline, focus on optimizing both your input documents and your prompts. Think about how a human expert would search this material.

• Structured Inputs: Upload clean, well-formatted documents (PDFs or TXT files) rather than complex layouts with multiple columns, which can confuse the chunking process.
• Specific Queries: Instead of asking 'Tell me about the project,' try 'What were the budget constraints mentioned in Q3 regarding Project Phoenix?' Specificity helps the cosine search narrow down the correct context.
Test Edge Cases: Run queries that test boundaries, such as cross-referencing data from two different documents to ensure the retrieval system can link disparate pieces of information.

The better your source material and query precision, the more reliable the top-5 context will be.