Rivet

Rivet is an IDE designed for creating complex AI agents and prompt chaining. It provides both a desktop application for visual graph creation and a core library for embedding those graphs into your own applications.

Key Capabilities:

• LLM Support: OpenAI (GPT-3.5, GPT-4), Anthropic (Claude Instant, Claude 2, Claude 3 Haiku, Sonnet, and Opus), and AssemblyAI LeMUR.
• Vector Database Support: OpenAI Embeddings and Pinecone.
• Integrations: Speech-to-Text from AssemblyAI.

Rivet is an IDE and library for building AI agents using a visual, graph-based interface. It is composed of two primary parts:

1. Rivet Application: A visual editor/IDE used to design complex prompt chains and AI agents. It includes tools like a prompt designer, A/B testing node variations, and integrated testing suites. The application produces project files that are designed for execution.

2. Rivet Core/Rivet Node: TypeScript libraries used to execute the project files created in the Rivet Application. These libraries allow you to treat your visual AI graphs as executable functions within your own software.

@ironclad/rivet-node is a Node.js binding for Rivet core. It provides helper APIs specifically for loading Rivet graphs from the filesystem and executing them.

Because it re-exports all types from @ironclad/rivet-core, you can use @ironclad/rivet-node as a drop-in replacement for the core package in Node.js environments.

Rivet provides several built-in plugins for integrating with external AI services, data providers, and evaluation tools. The available built-in plugins are:

• Anthropic: Integration with Anthropic's models.
• AssemblyAI: Integration with AssemblyAI for audio processing.
• Autoevals: Tools for automated evaluations.
• Gentrace: Integration with Gentrace.
• Google: Integration with Google services.
• HuggingFace: Integration with HuggingFace models and datasets.
• Pinecone: Integration with Pinecone vector database.

Refer to the specific documentation for each plugin to learn about its configuration, required API keys, and node usage.

The Prompt Node creates a chat-message object, which consists of a text string, a message type (User, Assistant, or System), and an optional name. It supports dynamic text interpolation using {{variable_name}} syntax, similar to a Text Node.

Key Capabilities

• Interpolation: Dynamically insert values into the prompt text by creating custom input ports based on the names used in the {{}} placeholders.
• Token Counting: If Compute Token Count is enabled in the editor settings, the node provides a Token Count output (number) representing the size of the generated message.
• Function Call Attachment: If Enable Function Call is enabled, you can attach a function call object to the chat message via a Function Call input port.

The Rivet CLI is built on top of @ironclad/rivet-node and provides two primary ways to interact with Rivet graphs:

1. Running graphs directly from the command line.
2. Starting a local HTTP server to run graphs via HTTP requests.

To use a local LLM provider like LM Studio instead of OpenAI:

1. Add a Chat node to your graph.
2. Add a Text node and place your message inside it (e.g., replacing {{input}} with your message).
3. Connect the output of the Text node to the Prompt input of the Chat node.
4. In the Chat node settings, set the Endpoint to http://localhost:1234/v1/chat/completions.
5. Load your model in LM Studio and ensure CORS is enabled in the LM Studio Server Options.
6. Run your graph.