[moekatib]

Pica's OneTool integrates with the Vercel AI SDK and LangChain, simplifying API connections for your LLM. It enhances AI capabilities without expanding the context window, reducing tool-related hallucinations. For example, to work with 40 tools, you'd typically need 4 agents and a router agent. With Pica, a single agent can access all 40 tools simply by adding OneTool to your LLM.

See how it works with Pica here https://github.com/picahq/onetool-demo

[lgas]

I'm guessing it's basically doing RAG over the full toolset to surface only the most relevant tools for any given message?

[moekatib]

It works very similarly to RAG but for tools. The SDK loads the necessary configs and schemas for each tool and uses that to guide the AI on how to use the tool effectively based on the latest message.