Descriptions:
Cole Medin demonstrates how to build a hybrid search RAG agent using MongoDB, Pydantic AI, and Dockling — a stack he argues is both simple and robust across a wide range of retrieval tasks. Hybrid search runs vector (semantic) search and BM25 keyword search simultaneously on every query, combining the ability to connect conceptually related content with the precision to match exact terms, identifiers, and named entities.
MongoDB serves as both the document store and vector database, handling chunk embeddings alongside full document records. Pydantic AI (version 1.27) powers the agent, with a retrieval tool that accepts a search_type parameter — hybrid, semantic, or keyword — letting the model choose the optimal strategy per query. Dockling processes ingestion from PDFs, Word documents, markdown, and audio files, and handles hybrid chunking to produce clean, well-bounded text segments for embedding.
Live demos illustrate where each strategy excels: a semantic query about a product launch timeline successfully retrieves meeting notes that never use the word “timeline” by mapping the concept to launch preparation content, while keyword search handles precise lookups by date or product name. Medin frames hybrid search as a form of agentic RAG — giving the agent agency over how it explores a knowledge base — and provides a GitHub template as a starting point for anyone building document retrieval systems.
📺 Source: Cole Medin · Published December 11, 2025
🏷️ Format: Tutorial Demo
