semantic-router

Why we included this project

If your agent or chatbot has to decide what to do with a given request, this library makes that call without burning a full LLM generation. You define routes as small sets of example queries, and incoming text gets embedded and matched against them, so a message can land as politics, chitchat, or a support ticket, whichever buckets you set up. Because the comparison happens in vector space rather than through a model generation, responses come back in milliseconds, and a fully local setup works with HuggingFace encoders and a small local LLM. It integrates with LangChain agents, handles multimodal input like image classification, and can route parameter generation and function calls through dynamic routes. The API is small enough to adopt in an afternoon, and the docs carry practical notebooks for threshold tuning plus real-world examples like healthcare admin workflows.