🔬 Autonomous Research Agent

A fully autonomous research agent built with OpenAI-Compatible LLMs + LangGraph that:

• Plans research tasks automatically
• Searches the web (Tavily / DuckDuckGo)
• Retrieves from uploaded documents (RAG via ChromaDB)
• Summarises content with map-reduce
• Maintains memory across tasks
• Generates structured Markdown / HTML reports

All free resources — can run entirely locally using Ollama or using free tier cloud APIs like Groq.

Example Output

---

Architecture

User Query
    │
    ▼
┌─────────────┐
│   Planner   │  LLM decomposes query → ordered task list
└──────┬──────┘
       │
       ▼
┌─────────────┐         ┌───────────────┐   ┌──────────────┐
│  Executor   │◄────────│  Web Search   │   │ Doc Retrieval│
│  (ReAct     │────────►│ Tavily / DDG  │   │ ChromaDB RAG │
│   loop)     │         └───────────────┘   └──────────────┘
└──────┬──────┘         ┌───────────────┐   ┌──────────────┐
       │                │  Summariser   │   │    Memory    │
       │                │  (Map-Reduce) │   │  (ChromaDB)  │
       ▼                └───────────────┘   └──────────────┘
┌─────────────┐
│  Summariser │  Section-level synthesis
└──────┬──────┘
       ▼
┌─────────────┐
│  Reporter   │  Jinja2 → Markdown + HTML report
└─────────────┘

---

Quick Start

1 — Clone & install

git clone <this-repo>
cd autonomous_research_agent
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2 — Configure

cp .env.example .env
# Open .env and set LLM_API_KEY and LLM_BASE_URL
# Optionally set TAVILY_API_KEY for better search quality

3a — Streamlit UI

streamlit run app.py
# Open http://localhost:8501

3b — CLI

# Simple query
python main.py "What are the latest breakthroughs in quantum computing?"

# With local documents
python main.py "Summarise the key arguments in the attached papers" \
    --docs paper1.pdf notes.txt

# Save report to custom path
python main.py "AI regulation in the EU" --output my_report.md

---

Project Structure

autonomous_research_agent/
├── agents/
│   ├── state.py          # ResearchState TypedDict
│   ├── planner.py        # Planner LangGraph node
│   ├── executor.py       # Executor ReAct loop node
│   └── graph.py          # Full LangGraph workflow
├── tools/
│   ├── web_search.py     # Tavily + DuckDuckGo
│   ├── doc_loader.py     # PDF / URL / text ingestion + chunking
│   └── summarizer.py     # Map-Reduce summarisation
├── memory/
│   ├── vector_store.py   # ChromaDB + sentence-transformers
│   └── agent_memory.py   # Agent memory add/query layer
├── report/
│   ├── generator.py      # Jinja2 report rendering + save
│   └── templates/
│       └── report_template.md
├── data/                 # Auto-created at runtime
│   ├── chroma_db/        # Persistent vector store
│   ├── reports/          # Generated reports
│   └── uploads/          # Saved uploads
├── app.py                # Streamlit UI
├── main.py               # CLI entry point
├── config.py             # Centralised settings
├── requirements.txt
└── .env.example

---

Free Resources Used

Component	Tool	Cost
LLM	Generic OpenAI / Groq / Ollama	Free / Low Cost
Web Search	DuckDuckGo	Free
Web Search+	Tavily API	1K/mo free
Vector DB	ChromaDB (local)	Free
Embeddings	sentence-transformers	Free (local)
Agent framework	LangGraph (OSS)	Free
RAG pipeline	Custom (LlamaIndex-free)	Free
Report	Jinja2 + Markdown	Free
UI	Streamlit	Free

---

Configuration Reference

Variable	Default	Description
LLM_API_KEY	empty	Your API key (optional if using Ollama)
LLM_BASE_URL	empty	Base URL (e.g., https://api.groq.com/openai/v1)
TAVILY_API_KEY	empty (uses DuckDuckGo)	Optional – better search
PLANNER_MODEL	llama-3.1-8b-instant	Model for planning
EXECUTOR_MODEL	llama-3.1-8b-instant	Model for execution
REPORTER_MODEL	llama-3.1-8b-instant	Model for report generation
MAX_TASKS_PER_PLAN	8	Max sub-tasks per research run
MAX_SEARCH_RESULTS	5	Results per web search
CHUNK_SIZE	500	Words per document chunk
CHUNK_OVERLAP	50	Overlap between chunks

---

Extending the Agent

Add a new tool

1. Create tools/my_tool.py with a function that returns str.
2. In agents/executor.py, add a new elif task["type"] == "my_tool": branch.
3. In agents/planner.py, add "my_tool" to the valid types in the system prompt.

Change the LLM

Set PLANNER_MODEL, EXECUTOR_MODEL, or REPORTER_MODEL in .env to any supported model string (e.g. llama-3.3-70b-versatile for Groq, or llama3 for Ollama).

Use a remote vector database

Replace the chromadb.PersistentClient in memory/vector_store.py with chromadb.HttpClient(host=..., port=...) or swap in Pinecone / Weaviate.

---

License

MIT — use freely.