Deep Research AI Agent

Generate comprehensive research reports on any topic using AI-powered web search and synthesis.

Problem

Researching a topic thoroughly is time-consuming. You need to find sources, evaluate credibility, synthesize information, and format citations—all before you even start writing.

Solution

This tool automates the research workflow:

1. Search — Queries Tavily API for relevant sources (filters out social media noise), with parallel variant queries and cross-encoder reranking
2. Cache — Stores results in ChromaDB with semantic search and smart TTL
3. Extract — Distills raw sources into a structured evidence table that grounds section prompts
4. Synthesize — LLM drafts structured reports in your chosen style; only failed sections are retried, not the whole report
5. Export — Download as PDF, Word, Markdown, or BibTeX

Live Demo →

Architecture

flowchart LR
    subgraph Input
        Q[Query]
        C[Config]
    end

    subgraph LangGraph["LangGraph Workflow"]
        R[Research Node]
        E[Evidence Extraction]
        D[Draft Node<br/>section-level retry]
    end

    subgraph Storage
        V[(ChromaDB<br/>Vector Store)]
        CA[(Joblib Cache)]
    end

    subgraph External
        T[Tavily API]
        O[OpenRouter LLM]
    end

    Q --> R
    C --> D
    R --> |search| T
    R --> |check cache| V
    R --> |store| V
    R --> |cache query| CA
    R --> E
    E --> |evidence table| D
    D --> |generate| O
    D --> Export[PDF/Word/MD/BibTeX]

Key Design Decisions:

• Two-node state machine: Research and Draft are decoupled for testability
• Vector memory with TTL: News content expires in 3 days; evergreen in 30 days
• Cross-encoder reranking: Improves retrieval precision over raw similarity
• Evidence table: Sources are distilled into a structured evidence table that grounds section prompts and inline [n] citations (unsupported citations are dropped)
• Parallel sub-queries: Deep-research variant queries run concurrently via ThreadPoolExecutor
• Section-level draft retry: A validation failure regenerates only the failed section, not the whole report
• Model fallback chain: On rate limits or empty responses, drafting moves down OPENROUTER_FALLBACK_MODELS and stays there once a fallback works
• Domain filtering: Excludes Reddit, Twitter, TikTok by default

Quick Start

# Clone
git clone https://github.com/saksham-jain177/AI-Agent-based-Deep-Research.git
cd AI-Agent-based-Deep-Research

# Install
pip install -r requirements.txt

# Configure
cp .env.example .env
# Edit .env with your API keys

# Run
streamlit run app.py

Required API Keys:

Optional:

ENABLE_VECTOR_STORE=true       # Enable ChromaDB caching
PREFER_CACHE_RESULTS=false     # Prefer cached over fresh results
LLM_MAX_OUTPUT_TOKENS=4000     # Per-call LLM output token cap (default 4000)
TOKEN_BUDGET_PER_REPORT=12000  # Total prompt-token budget; low-confidence evidence is trimmed to fit
OPENROUTER_FALLBACK_MODELS=    # Comma-separated fallback model chain for rate limits / empty responses

Configuration

Performance

Measured on typical queries with ChromaDB caching enabled.

Non-Goals

This tool is not:

• A fact-verification engine — always verify critical claims
• Real-time streaming — uses section-by-section progress
• A citation authority — check sources before academic submission
• Optimized for legal/medical/financial advice

Limitations

• Rate limits: Tavily free tier has daily limits
• Model availability: OpenRouter free models may be rate-limited (a fallback model chain mitigates this)
• Citation accuracy: Auto-generated citations should be manually verified; inline [n] references are checked against the evidence table, but underlying source quality is not judged
• Language support: Best results in English; ES/DE are functional but less tested

Testing

The test suite (180 tests) runs fully offline — all LLM and Tavily calls are mocked. It covers evidence extraction, citation validation, section-level draft retry, reranking, prompt-injection sanitization, log redaction, token budgets, output-token caps, model fallback, and timeouts.

pytest tests/ -v

Project Structure

├── app.py                # Streamlit UI
├── main.py               # LangGraph workflow orchestration
├── research_agent.py     # Tavily search (parallel variants) + vector store integration
├── evidence_extractor.py # LLM-assisted distillation of sources into an evidence table
├── draft_agent.py        # LLM prompting, section generation, retries, fallback chain, cost controls
├── vector_store.py       # ChromaDB with cross-encoder reranking and TTL
├── sanitize.py           # Prompt-injection neutralization for fetched web content
├── log_redaction.py      # API key redaction from logs and session state
├── cost_estimator.py     # Token/cost estimation with uncertainty
├── citation_formatter.py # APA/MLA/IEEE/BibTeX generation with citation grounding
└── tests/                # Pytest suite: 180 tests, all offline (LLM/Tavily mocked)

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/your-idea)
3. Run tests (pytest tests/ -v)
4. Submit a pull request

See CONTRIBUTING.md for detailed guidelines.

License

MIT License — use freely for personal or commercial projects.

Questions? Open an issue or start a discussion.