Deep Research Skill

An open-source multi-hop research, fact verification & anti-hallucination engine for AI agents and automation pipelines.

---

The Problem with AI Research Today

Standard LLMs fail at deep, reliable research:

• ❌ Hallucinated Citations — Citing broken URLs, fake paper titles, or made-up statistics.
• ❌ Superficial Summaries — A 2-paragraph surface-level output instead of a deep multi-angle investigation.
• ❌ Single-Query Blindspots — Relying on a single search vector instead of structured, recursive research trees.

deep-research-skill solves all three problems with an open-source, multi-hop recursive research engine that scrapes primary sources, cross-audits claims across multiple web domains, calculates statistical confidence scores, and outputs zero-hallucination research reports with verified inline citations.

---

System Architecture

             +-----------------------------+
             |   User Research Topic       |
             +-------------+---------------+
                           |
                           v
             +-----------------------------+
             |   Multi-Hop Search Planner  |
             |  (Overview / Technical /    |
             |   Verify / Counter-View)    |
             +-------------+---------------+
                           |
                           v
             +-----------------------------+
             |  Async Web Scraper          |
             |  HTML → LLM Clean Markdown  |
             +-------------+---------------+
                           |
                           v
             +-----------------------------+
             |  Cross-Source Fact Auditor  |
             |  Confidence Score 0–100%    |
             +-------------+---------------+
                           |
                           v
             +-----------------------------+
             |  Report Synthesizer         |
             |  Executive MD + [[1]] Cites |
             +-----------------------------+

---

Key Features

• Multi-Hop Search Planner — Breaks research topics into 4 orthogonal query vectors: Overview, Technical Specs, Primary Verification, and Counter-Perspectives.
• Zero-Clutter Web Scraper — Strips ads, navigation bars, and scripts to feed clean, dense text to any LLM.
• Cross-Source Fact Auditor — Cross-checks factual assertions across multiple distinct web domains and computes an Anti-Hallucination Confidence Score (0–100%).
• Verified Citation Indexing — Auto-generates verified inline footnote citations [[1]](url) in all reports.
• Universal AI Compatibility — Works with any AI system: LIGHT LPCA, OpenAI, Claude, Groq, Ollama, or standalone scripts. No lock-in.

---

30-Second Quickstart

Install

pip install deep-research-skill

Or clone for development:

git clone https://github.com/LIGHTLPCA/deep-research-skill.git
cd deep-research-skill
pip install -e .[dev]

Run via CLI

deep-research "Large Language Model Hallucination Benchmarks 2026" --out report.md

Python API

from deep_research import DeepResearchAgent

agent = DeepResearchAgent(max_queries_per_dimension=1, max_pages_per_query=2)
report = agent.run("Quantum Computing Error Correction")

print(f"Confidence Score : {report.confidence_score}%")
print(f"Verified Claims  : {report.verified_claim_count}")
print(report.markdown_content)

---

Sample Report Output

# Deep Research Report: Quantum Computing Error Correction

**Generated by Deep Research Skill**
> Anti-Hallucination Confidence Score: 91.4%
> Total Pages Analyzed: 8 | Verified Claims: 6

## Verified Core Findings & Audit Trail

### 1. Logical Qubit Fidelity Threshold Exceeds 99.9%
- **Status:** `VERIFIED` (Confidence: `94%`)
- **Verified Sources:** [[1]](https://example.org/qc-paper) [[2]](https://example.org/qc-benchmarks)

## Primary Citation Index
1. https://example.org/qc-paper
2. https://example.org/qc-benchmarks

---

Running Tests

pytest tests/

---

Contributing

We welcome community contributions! Please read CONTRIBUTING.md for setup instructions.

Open areas for contribution:

• New search API providers (Tavily, SearXNG, Brave)
• PDF & HTML export support
• ArXiv academic paper parser
• Rich terminal UI dashboard

See ROADMAP.md for detailed milestones.

---

License

Distributed under the MIT License. See LICENSE for details.