@varlabs/ai.signatures

DSPy/ax-style typed signatures and a ReAct agent loop for the AI SDK: a typed input/output spec that compiles into a prompt and parses the model's response back into the declared shape - no hand-written prompt strings.

Installation

npm install @varlabs/ai.signatures
# or
yarn add @varlabs/ai.signatures
# or
pnpm add @varlabs/ai.signatures

Usage

Signatures

import { defineSignature } from '@varlabs/ai.signatures';

const qa = defineSignature({
  input: { question: { type: 'string' } },
  output: { answer: { type: 'string' } },
  instructions: 'Answer the question concisely.',
});

const prompt = qa.compilePrompt({ question: 'What is the capital of France?' });
const modelResponse = await myModelCall(prompt);
const { answer } = qa.parseOutput(modelResponse);

ReAct agent

runReActAgent loops a signature plus a tool set until the model produces the signature's final output (or maxSteps is exceeded). Server-located tools run inline; client-located tools are handed to onClientToolCall.

import { runReActAgent } from '@varlabs/ai.signatures';

const result = await runReActAgent(
  {
    signature: qa,
    tools: { search: mySearchTool },
    callModel: (prompt) => myModelCall(prompt),
  },
  { question: 'What is the capital of France?' },
);

License

MIT