Descriptions:
Cole Medin walks through building a RAG agent with real-time source validation and human-in-the-loop chunk approval, directly addressing one of the most persistent failure modes in production RAG systems: not knowing whether the AI actually used the right documents — or fabricated citations entirely.
The architecture pairs CopilotKit on the frontend with a Pydantic AI agent on the backend, connected via the AGUI protocol. When a user submits a query, the agent retrieves chunks from the knowledge base and surfaces them in the UI for manual review before generating any answer. The key enabler is the useAgent hook, recently released by the CopilotKit team, which provides real-time bidirectional state sync between the React frontend and the backend agent. Users can select or deselect individual chunks, and the agent immediately reflects that curated context when synthesizing its response.
Medin also demonstrates using this setup as a RAG debugging harness — plugging any existing RAG agent into the frontend to verify what knowledge it’s actually drawing on during development. The video includes an Excalidraw architecture diagram and a code walkthrough covering state snapshot events, dependency injection in Pydantic AI, and the AGUI endpoint configuration. Developers building explainable or high-stakes RAG applications will find this a practical reference for implementing controlled, auditable retrieval pipelines.
📺 Source: Cole Medin · Published December 22, 2025
🏷️ Format: Hands On Build
