Introducing JigsawStack Deep Research

What is Deep Research?

JigsawStack’s Deep Research is an open source framework performing for multi-hop, AI-assisted research. The framework orchestrates large language models (LLMs), recursive web searches, and structured reasoning to generate comprehensive, evidence-backed reports via our deep-research SDK. Designed for developers, researchers, and AI agents. We aim to automate the kind of deep inquiry that typically takes hours or even days.

Get a quick start by trying out our SDK here: https://github.com/JigsawStack/deep-research

import { createDeepResearch } from "deep-research";

const deepResearch = createDeepResearch({
  OPENAI_API_KEY: process.env.OPENAI_API_KEY,
  GEMINI_API_KEY: process.env.GEMINI_API_KEY,
  DEEPINFRA_API_KEY: process.env.DEEPINFRA_API_KEY,
  JIGSAW_API_KEY: process.env.JIGSAW_API_KEY,
});

const result = await deepResearch.generate("What are the recent developments in quantum computing?");

console.log(result.data.text);           // Full research report
console.log(result.data.bibliography);   // Auto-generated references

Core Concepts

How JigsawStack Implements Deep Research

Deep Research performs the following:

1. Research Plan, decomposing complex query into logical subtopics and objectives.

2. AI Web Searches via JigsawStack API, generating context and extracting key points from the web

3. Multi-Model Reasoning and Decision Making, the framework adapts its strategy based on findings, filling knowledge gaps as needed

4. Report Generation, including structured content and source attribution for traceability

Fully Customizable

One of the core strengths of JigsawStack’s Deep Research is how deeply customizable it is.

• Precise Control Over Depth & Breadth

• Report Length Targeting - Configure your desired report length with soft and hard output limits

• Developer-Friendly Debugging - Enable detailed logging for developers

• Model Modularity - Mix and match AI Providers based on their strengths

const deepResearch = createDeepResearch({
  report: {
    targetOutputTokens: 10000,  // Target length for the generated report
    maxOutputTokens: 30000,     // Hard ceiling for total output
  },
  depth: { maxDepth: 4 },       // How many recursive research cycles to run
  breadth: { maxBreadth: 3 },   // How many subqueries to generate per level
  models: {
    default: openaiModel,       // Main LLM for prompt interpretation
    reasoning: deepseekModel,   // Specialized model for analyzing research findings
    output: geminiModel,        // Model used for final report generation
  },
  logging: { enabled: true },   // Enable verbose logging for debugging
});

Current Market and the Challenges of Deep Research

While many are pushing boundaries, current implementations face a few common challenges:

• Citation Accuracy & Transparency

• Limited Customization & Developer Control

• Closed-Source & Opaque Systems

• Limited Integration Options

We aim to address these pain points, making it a research platform you can build on.

Comparisons

Real World Examples

Write a research paper talking about diffusion for LLMs and how it might affect the future in technology?

const query = "Write a research paper talking about diffusion for LLMs and how it might affect the future in technology?";

const result = await deepResearch.generate(query);

console.log(result.data.text) // report
console.log(result.data.bibliography) // references

const singleReport = result.data.text + "\n\n" + result.data.bibliography

What is the largest order of a non-cyclic torsion subgroup of an elliptic curve over \mathbb{Q}(\sqrt{-3})\)

const query = "What is the largest order of a non-cyclic torsion subgroup of an elliptic curve over \mathbb{Q}(\sqrt{-3})";

const result = await deepResearch.generate(query);

console.log(result.data.text) // report
console.log(result.data.bibliography) // references

Improvements to JigsawStack’s Deep Research

While Deep Research provides powerful automation for complex inquiry, it’s not without its trade-offs and design challenges:

• Runtime Trade-offs - deep research deliberately shifts away from the fast-but-shallow answers towards deep and completed ones.

• Exclusive to Public Information - we are unable to access paywalled research papers, private documents, or internal knowledge bases

• Sources Discrepancy - Some sources provided are not quoted on the report. Even if a URL isn’t directly cited in the final report, it may still play a critical role in shaping the system’s reasoning.

• Evolving output format - The first stage of Deep Research prioritizes correctness, structure, and citations. But the ideal report format — tone, length, granularity — will vary by user and use case. We’re actively evolving the output schema based on real-world feedback to support!

Next Steps

As we continue to evolve Deep Research, the next major milestone is full integration into JigsawStack’s web search API.

Currently, Deep Research uses JigsawStack’s search API for retrieving public content. Going forward, we’re expanding this connection to make the research experience even more seamless and intelligent.