Google Embedding UTCP Manual

This project provides a UTCP (Universal Tool Calling Protocol) manual for the Google Gemini embedding model. This manual allows any UTCP-compliant client to generate text embeddings using the Gemini API without requiring any custom wrapper code.

About the Google Embedding Model

Text embeddings are numerical representations of text that capture semantic meaning. These vector representations are crucial for a wide variety of NLP tasks, including:

• Semantic Search: Finding text that is conceptually similar, not just keyword-matched.
• Classification: Categorizing text into predefined labels.
• Clustering: Grouping similar texts together to identify trends.
• Retrieval Augmented Generation (RAG): Enhancing language model outputs with factual information retrieved from a knowledge base.

The Google Gemini embedding model (gemini-embedding-001) is a state-of-the-art model that generates high-dimensional embeddings (3072 dimensions by default) to power these applications.

About the Universal Tool Calling Protocol (UTCP)

UTCP is a modern, flexible, and scalable standard for defining and interacting with tools across various communication protocols. Its core philosophy is to describe, not wrap.

Instead of building and maintaining custom adapter servers for every tool, UTCP allows developers to create a simple "manual" (a .json file) that teaches a client how to call the tool's native endpoint directly. This approach has several key benefits:

• No Wrapper Tax: Eliminates the overhead of building and deploying adapter services.
• Direct Communication: Clients call tools directly, reducing latency.
• Leverages Existing Infrastructure: Uses the tool's native authentication, security, and rate-limiting.
• Protocol Agnostic: Works with any communication protocol (HTTP, WebSockets, CLI, gRPC, etc.).

Usage

To use this manual, you will need a UTCP client. The client will load the google-embedding-utcp-manual.utcp.json file to discover the embed_content tool.

Calling the embed_content Tool

Here is an example of how a client would call the embed_content tool defined in this manual:

{
  "tool": "embed_content",
  "inputs": {
    "model": "gemini-embedding-001",
    "contents": ["Hello world"]
  }
}

Authentication

The tool_provider in the google-embedding-utcp-manual.utcp.json file uses a variable for the API key: ${GEMINI_API_KEY}. Your UTCP client must be configured to replace this variable with your actual Gemini API key.