Pi Hyper-Search 🚀

An open-source, ultra-fast, highly parallel search plugin for the Pi Agent Runtime.

Features

• Dynamic Intensity: Exposes an intensity parameter (1-5) to the main LLM. Scales from 5 parallel queries up to 30 parallel queries and 50 URLs fetched in seconds.
• Asynchronous Multiplexed Swarm: Fires parallel Serper queries using native array batching.
• In-flight Edge Filtering: Fetches top URLs concurrently using Spider's high-speed HTTP mode, with strict early termination timeouts.
• Smart Model Routing: Integrates with Pi's ModelRegistry to automatically select the fastest available model (e.g. qwen-flash, gemini-3.1-flash-lite) for the query fan-out phase.
• Self-Healing Retries: Automatically monitors scraped text volume and re-prompts for different queries if the initial swarm returns insufficient data.

Installation

npm install Fornace/pi-hypersearch

Usage

Register the tool in your Pi runtime:

import { createHyperSearchTool } from 'pi-hypersearch';

const hyperSearch = createHyperSearchTool({
  token: process.env.FC_AUTH_TOKEN, // If using Fornace Proxy
  fornaceHost: 'https://fornace.net',
  // OR provide direct keys:
  // serperApiKey: process.env.SERPER_API_KEY,
  // spiderApiKey: process.env.SPIDER_API_KEY,
  // openAiApiKey: process.env.OPENAI_API_KEY,
  modelRegistry: myModelRegistry, // For smart routing
});

// Add to your customTools array

How It Works

1. Flash Decomposition: A fast model breaks the user's complex prompt into multiple specific Google queries.
2. Batch Search: All queries are sent to Serper in a single batched HTTP request.
3. Concurrent Scrape: The top unique URLs are scraped in parallel using Spider in HTTP mode to bypass Javascript rendering overhead.
4. Synthesis: The dense Markdown is returned to the main Pi agent.