# 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](https://github.com/JigsawStack/deep-research)

```typescript
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

| **Concept** | **Description** |
| --- | --- |
| **Deep Thinking** | The system breaks down a question into logical parts, reasons through them independently, and synthesizes an answer. |
| **Deep Research** | The system performs multi-hop, focused web searches, compares the findings, and composes an evidence-backed answer |

## 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
    

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1748474234103/62296313-aaf4-4e3f-8cfd-122f875ce5a7.png align="center")

## 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
    

```typescript
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

|  | **JigsawStack** | **OpenAI** | **Perplexity** |
| --- | --- | --- | --- |
| Citation Transparency | ✅ Strong. Ensures source traceability with comprehensive bibliographies. | ⚠️ Partial. While citations are provided, some users report occasional inaccuracies or reliance on less authoritative sources. | ✅ Strong. Curates a list of relevant URLs with titles and descriptions for each research task. |
| Customization & Control | ✅ Full. Allows developers to configure every stage, including depth, models, and formatting. | ❌ Limited. Users have minimal control over the research process and output formatting. | ⚠️ Limited. Offers some customization, such as model selection, but lacks deep configurability. |
| Open Source | ✅ Yes. Fully open-source. | ❌ No. | ❌ No. |
| API / SDK Access | ✅ Yes. accessible SDK for developers to plug and play into their workflow. | ⚠️ Internal use only | ⚠️ Limited / closed |
| Limitations | Slower runtime by design; limited to public content only. | May produce inconsistent information and occasionally cite less authoritative sources. | Lacks creative flexibility and may provide less engaging conversational responses. |

## Real World Examples

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

```typescript
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
```

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1748460611833/74a9e8f7-6255-4a0f-b2b9-9127e873e5fc.png align="center")

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

```typescript
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
```

![](https://cdn.hashnode.com/res/hashnode/image/upload/v1748461313544/9e061c0b-797b-49ee-a3f4-a29a0da4f03d.png align="center")

## 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.