🔭 Orion — Autonomous Research Agent

Drop a question. Get a cited, structured research report in minutes.

---

---

🌐 Live Demo

🚀 Try Orion Live - No installation required!

---

📸 Screenshots

Title & Branding

Dashboard

Search & Query

Research Output

---

🧠 How It Works

Orion is a multi-agent research pipeline. You ask a question — it autonomously plans, searches, remembers, synthesizes, and critiques until the report meets quality standards.

Your Query
    │
    ▼
🧠 Planner (Llama 3.3 70B)     → breaks into targeted sub-questions
    │               │
    ▼               ▼
🌐 Tavily Search   💾 FAISS Recall    ← run in parallel
    │                   │
    ▼                   │
💾 FAISS Store ◄────────┘            ← persist new findings
    │
    ▼
✍️  Synthesizer (Llama 3.3 70B)      ← writes structured report
    │
    ▼
🔍 Critic (Llama 3.3 70B)           ← scores quality 0.0 → 1.0
    │               │
    ▼               ▼
  DONE ✅      loop back 🔁          ← iterate if score < 0.8

Agent nodes

Node	Role	Tool
planner	Decomposes query into sub-questions	Llama 3.3 70B (Groq)
searcher	Fetches live web evidence	Tavily
memory_retrieve	Recalls relevant past research	FAISS
memory_store	Embeds & persists new findings	FAISS
synthesizer	Compiles everything into a report	Llama 3.3 70B (Groq)
critic	Scores quality, loops or finishes	Llama 3.3 70B (Groq)

---

🚀 Quick Start

1. Clone & install

git clone https://github.com/YOUR_USERNAME/orion-research-agent
cd orion-research-agent
python -m venv .venv
.venv\Scripts\activate        # Windows
pip install -r requirements.txt

2. Get free API keys

Key	Where to get
GROQ_API_KEY	console.groq.com — free, no credit card
TAVILY_API_KEY	app.tavily.com — free tier: 1000 searches/month

3. Configure

copy .env.example .env    # Windows

Edit .env:

GROQ_API_KEY=gsk_xxxxxxxxxxxxxxxxxxxxxxxx
TAVILY_API_KEY=tvly-xxxxxxxxxxxxxxxx

4. Run

streamlit run app.py

Open http://localhost:8501 and start researching. 🔭

---

⚙️ Configuration

All parameters are in config.py:

Setting	Default	Description
GROQ_MODEL	llama-3.3-70b-versatile	LLM for all reasoning nodes
MAX_SUB_QUESTIONS	5	Sub-questions the planner generates
SEARCH_RESULTS_PER_QUERY	5	Tavily results per sub-question
MEMORY_TOP_K	5	FAISS chunks recalled per run
MAX_ITERATIONS	3	Research loop safety cap
QUALITY_THRESHOLD	0.8	Critic score (0–1) to accept report
EMBEDDING_MODEL	all-MiniLM-L6-v2	SentenceTransformer for FAISS

---

📁 Project Structure

Orion/
├── app.py                  ← Streamlit web UI
├── graph.py                ← LangGraph pipeline definition
├── state.py                ← AgentState TypedDict
├── config.py               ← All configuration & API keys
├── main.py                 ← CLI entry point
├── agent_nodes/
│   ├── __init__.py
│   ├── planner.py          ← Sub-question generation (Groq)
│   ├── searcher.py         ← Web search (Tavily)
│   ├── memory.py           ← FAISS store + retrieve
│   ├── synthesizer.py      ← Report writing (Groq)
│   └── critic.py           ← Quality evaluation (Groq)
├── screenshot/             ← UI screenshots
├── faiss_store/            ← Persistent vector index (auto-created)
├── reports/                ← Saved research reports (auto-created)
├── requirements.txt
└── .env.example

---

🔧 Extending Orion

Add a new node (e.g. PDF reader):

1. Create agent_nodes/pdf_reader.py with pdf_reader_node(state) -> dict
2. Register in graph.py: graph.add_node("pdf_reader", pdf_reader_node)
3. Wire it: graph.add_edge("searcher", "pdf_reader")

Swap the LLM — change GROQ_MODEL in config.py to any model on Groq's free tier.

Async search — replace sequential Tavily calls in searcher.py with asyncio.gather for faster parallel fetching.

---

👤 Author

Sumit Mishra

• GitHub: @sumittt2004
• LinkedIn: Sumit Mishra

---

Built with LangGraph · Groq · Tavily · FAISS · Streamlit