Ask Anything
Drop in any document, PDF, or URL. Then ask it anything. A live RAG pipeline - retrieval-augmented generation in your browser.
Overview
What is RAG?
Retrieval-Augmented Generation (RAG) is one of the most practically useful patterns in applied AI. Instead of relying pureply on what a language model was trained on, RAG grounds the model's responses in a specific document of knowledge source you provide - making answers accurate, specific, and verifiable.
This demo lets you experience that pipeline directly. Paste in a contract, upload a research paper, drop in a news article URL - then ask qustions ang get answers grounded in that exact content. No allucinations about things not in the document. No generic responses.
Under the hood, the document is passes as context to Claude via the Anthropic API, with a system propmt engineered to keep responses grounded, cite specific sections, and flag when something isn't in the document.
How It Works
You provide a document - plan text, a PDF file, or any publically accessible URL. The content is extracted and cleaned.
The document is passed as context in the system prompt, with instructions to ground all answers in the provided content.
When you ask a question, Claude retrieves the relevant parts of the context and generate a grounded, cited response.
The full conversation history is maintained so you can ask follow-up questions and dig deeper into teh document.
What I Learned
Prompt engineering matters enormously for RAG - the difference between a grounded response and a hallucinated on is often just the system prompt.
PDF text extraction is messier than expected - tables, headers, and multi-column layouts all need special handeling to produce clean context.
Conversation history management is critical - including too much history bloats the context window, too little loses coherence.
Previous Project
Cadence