OmniResearch

Autonomous Deep Research and Fact-Checked Intelligence Engine

Autonomous multi-agent system that searches the public web and private enterprise archives to generate verified, cited executive dossiers.

Features • Architecture • Quick Start • Interfaces • Verification • Repository Structure

---

Overview

Most LLM research tools fail in corporate settings for two reasons: they hallucinate numerical claims, and they cannot cross-reference public web intelligence with private internal documentation.

OmniResearch runs an autonomous multi-agent state machine that plans investigative tracks, collects evidence from hybrid sources, audits every critical factual assertion against primary text excerpts, and exports formatted dossiers to Markdown, PDF, and DOCX.

Features

• Multi-agent state machine. Divides research between four focused agents: Planner, Researcher, Verifier, and Writer.
• Zero-cost web search. Runs out of the box using DuckDuckGo search and clean web scraping without requiring paid search API keys. Supports Tavily and Exa adapters when configured.
• Hybrid retrieval. Ingests local corporate files (.pdf, .docx, .csv, .md) alongside live web queries.
• Automated claim verification. Matches numbers, statistics, and critical assertions against source snippets, assigning an empirical confidence score from 0 to 100%.
• Model-agnostic LLM client. Connects to Google Gemini, OpenAI, Anthropic Claude, or local Ollama instances (Llama 3.3, DeepSeek, Qwen) for air-gapped security.
• Multi-format deliverables. Exports structured dossiers to Markdown, Word (.docx), and PDF with numbered source citations.
• Dual interfaces. Provides a Streamlit web workstation and a Rich CLI for terminal workflows.

Architecture

flowchart TD
    User([User / Corporate Analyst]) --> Input{Input Gateway}
    Input -->|Web UI| WebUI[Streamlit Workspace]
    Input -->|Terminal| CLI[Typer + Rich CLI]

    subgraph Agent_Orchestrator [OmniResearch State Machine]
        Planner[1. Planner Agent]
        Planner -->|Decomposed Hypotheses| Researcher[2. Researcher Agent]

        subgraph Hybrid_Retrieval [Hybrid Retrieval Hub]
            DDG[DuckDuckGo Free Search]
            Tavily[Tavily / Exa API]
            DocReader[Internal Docs: PDF / DOCX / CSV]
        end

        Researcher <--> Hybrid_Retrieval
        Researcher -->|Raw Insights & Citations| Verifier[3. Fact-Checker Agent]
        Verifier -->|Cross-Source Match & Confidence Audit| Verifier
        Verifier -->|Verified Ground Truth| Writer[4. Executive Writer Agent]
    end

    WebUI <--> Agent_Orchestrator
    CLI <--> Agent_Orchestrator

    Writer --> Exporters[Export Engine]
    Exporters --> Markdown[Markdown Dossier]
    Exporters --> PDF[Executive PDF Report]
    Exporters --> DOCX[Corporate Word Document]

Loading

Quick Start

1. Clone and Setup

git clone https://github.com/davinmaritza/omniresearch.git
cd omniresearch

# Create virtual environment and install dependencies
uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"

2. Configure Environment

Copy the example environment file:

cp .env.example .env

Configure your preferred model provider in .env (Gemini, OpenAI, Anthropic, or local Ollama):

LLM_PROVIDER=gemini
GEMINI_API_KEY=your_gemini_api_key_here
SEARCH_PROVIDER=duckduckgo

If no API keys are provided, OmniResearch runs with its built-in offline heuristic engine and DuckDuckGo search so you can test the entire pipeline immediately.

Interfaces

Web Dashboard (Streamlit)

Launch the interactive workstation:

python -m omniresearch.cli web
# or: streamlit run src/omniresearch/web.py

The web dashboard provides drag-and-drop document upload, live agent status telemetry, a searchable fact-check table, and 1-click downloads for Markdown, PDF, and DOCX files.

Terminal CLI

Run deep research directly from the command line:

# Research a market topic
python -m omniresearch.cli run "AI Chip Market Competitors: Nvidia vs AMD vs Custom Silicon 2026"

# Research with private local documents
python -m omniresearch.cli run "Q3 Performance Review" --docs ./data/internal_q3.pdf --export md,pdf,docx --out ./reports

To audit system readiness and configured providers:

python -m omniresearch.cli check

Verification and Fact-Checking

Unverified figures are liabilities in business decisions. OmniResearch routes all candidate evidence through the Verifier Agent before synthesis:

1. Named entity and metric extraction. Isolates numerical claims, dates, and market metrics.
2. Ground-truth cross-matching. Traces claims back to raw snippet chunks or document pages.
3. Confidence score assignment:
   • High (>= 85%): Direct corroboration in primary evidence.
   • Moderate (70% - 84%): Inferred from secondary corroboration.
   • Unverified (< 70%): Flagged in the appendix table for manual analyst review.

Repository Structure

omniresearch/
├── src/
│   └── omniresearch/
│       ├── cli.py             # Typer + Rich CLI entrypoint
│       ├── web.py             # Streamlit visual dashboard
│       ├── config.py          # Configuration and settings schema
│       ├── core/
│       │   ├── state.py       # Pydantic state models and schemas
│       │   ├── graph.py       # Multi-agent state machine workflow
│       │   └── llm.py         # Multi-provider LLM client adapter
│       ├── agents/
│       │   ├── planner.py     # Sub-query decomposition agent
│       │   ├── researcher.py  # Hybrid web and document retrieval agent
│       │   ├── verifier.py    # Fact-checking and confidence scoring agent
│       │   └── writer.py      # Executive dossier synthesis agent
│       ├── tools/
│       │   ├── search.py      # DuckDuckGo, Tavily, and Exa search adapters
│       │   ├── document.py    # PDF, DOCX, and CSV document parsers
│       │   └── scraper.py     # HTML text extraction utility
│       └── exporters/
│           ├── markdown.py    # Markdown report exporter
│           ├── pdf.py         # ReportLab PDF generator
│           └── docx.py        # Word document generator
├── tests/
│   └── test_pipeline.py       # End-to-end integration test suite
├── docs/                      # Architecture and enterprise documentation
├── pyproject.toml             # Project manifest and dependencies
└── LICENSE                    # Apache-2.0 License

Testing

Run the test suite with pytest:

pytest -v

License

Apache License 2.0. See LICENSE for details.