Powered by Groq JigsawStack Prompt Engine is a powerful AI SDK designed to integrate into any backend, automating tasks such as web scraping, optical character recognition (OCR), translation, and more, using custom fine-tuned models. By plugging JigsawStack into your existing application infrastructure, you can offload the heavy lifting and focus on building. The JigsawStack Prompt Engine is powered by the Groq LPU Inference Engine to deliver an exceptionally efficient workflow optimized for real-time performance. By leveraging Groq in the backend for fast inference and low latency, the JigsawStack Prompt Engine provides rapid results for your AI applications. Packed with a range of built-in features, Prompt Engine makes working with LLMs effortless and efficient:
  • 🌐 Prompt caching for repeated prompt runs
  • 💬 Automatic prompt optimization for improved performance
  • 📄 Response schema validation for accuracy and consistency
  • 🔁 Reusable prompts to streamline your workflow
  • 🧠 Multi-agent LLM from 50+ models for flexibility depending on your apps
  • 🚫 No virtual rate limits, tokens, and GPU management

​
How does it work?

JigsawStack Prompt Engine workflow The JigsawStack Prompt Engine is based on a Mixture-of-Agents (MoA) approach. Each time a prompt is executed, it is run across 5 LLMs under the hood, including LLMs powered by Groq for lightning-fast inference speed and performance, such as:
  • llama-3.3-70b-versatile
  • llama-3.1-8b-instant
  • other latest models from Groq
The output of each LLM is then ranked by a smaller model based on similarity and quality before being merged into a single output. The JigsawStack Prompt Engine especially works well if you run the same base prompt repeatedly to allow the engine to self-tune for running the best model every single time with the built-in prompt caching feature. Optionally each prompt execution is first ran through llama-guard-4-12b to detect and filter for common abuse such as prompt injection, crimes, sexual content and more. Prompt guard is a feature that’s built into every step of the Prompt Engine and is configurable to allow or block specific content types.
Learn more about how prompt engine works here

​
Getting Started

​
Prerequisite

​
Installation

npm i jigsawstack

​
Usage

You can get a JigsawStack API key by creating an account here. Store the key in a secure environment like a .env file.

​
Creating a prompt engine:

import { JigsawStack } from "jigsawstack";

const jigsaw = JigsawStack({ apiKey: "your-api-key" });

const result = await jigsaw.prompt_engine.create({
  prompt: "How to cook {dish}",
  inputs: [{ key: "dish" }],
  return_prompt: [
    {
      step: "step counter",
      instructions: "details of this step",
    },
  ],
});
All prompts are stored allowing you to rerun them with just the prompt ID

​
Running the prompt:

const resp = await jigsaw.prompt_engine.run({
  id: result.prompt_engine_id,
  input_values: {
    dish: "Singaporean chicken rice",
  },
});

​
Running a prompt directly:

const result = await jigsaw.prompt_engine.run_prompt_direct({
  prompt: "Tell me a story about {about}",
  inputs: [
    {
      key: "about",
      optional: true,
      initial_value: "Leaning Tower of Pisa",
    },
  ],
  return_prompt: "Return the result in a markdown format",
  input_values: {
    about: "Santorini",
  },
});
You can learn more about the Prompt Engine in the docs

​
Llama Guard 4 by Groq

The prompt engine comes with prompt guards to prevent prompt injection from user inputs and a wide range of unsafe use cases. This can be turned on automatically using the prompt_guard field. 
const result = await jigsaw.prompt_engine.run_prompt_direct({
  prompt: "Tell me a story about {about}",
  inputs: [
    {
      key: "about",
      optional: true,
      initial_value: "Leaning Tower of Pisa",
    },
  ],
  return_prompt: "Return the result in a markdown format",
  prompt_guard: ["sexual_content", "defamation"],
  input_values: {
    about: "Santorini",
  },
});

​
Prompt Guard Categories

  • defamation: Guards against potentially libelous content
  • privacy: Protects personal information
  • hate: Filters out discriminatory content
  • sexual_content: Blocks explicit material
  • elections: Prevents election misinformation
  • code_interpreter_abuse: Guards against code exploitation
  • indiscriminate_weapons: Blocks content about mass destruction weapons
  • specialized_advice: Prevents unauthorized professional advice
By implementing prompt guard, you can create more secure, reliable, and user-friendly AI-powered applications. It’s an invaluable tool for developers looking to harness the power of AI while maintaining control over the content generated and processed by their applications.
For the most up-to-date list of prompt guard options and detailed usage, refer to the full documentation.