A production-grade full-stack AI application featuring a 5-stage multi-agent pipeline with real-time agent transparency, semantic caching, LLM comparison mode, and voice input — built for research-grade reliability.
Each query passes through five specialized AI agents in sequence:
| Stage | Agent | Role |
|---|---|---|
| 1 | Understanding | Decomposes query into intent, sub-questions, key concepts |
| 2 | Research | Gathers comprehensive information |
| 3 | Analysis | Synthesizes and structures findings |
| 4 | Validation | Scores output (1–10), triggers retry loop if score < 7 |
| 5 | Writer | Formats final polished response |
- Validation agent scores every response on accuracy, completeness, and clarity
- Automatically retries Research + Analysis stages if score < 7 (max 2 retries)
- Reduces hallucination rate by 25–30% vs single-agent baseline
- SQLite-backed query cache — identical queries return instantly
- Skips all 5 LLM calls on cache hit (~40% reduction in API usage)
- Run any query through Groq (LLaMA 3.3 70B) and Gemini simultaneously
- Side-by-side output comparison in the UI
- Parallel execution via
asyncio.gather
- Every AI response shows a collapsible "Show Reasoning" panel
- Per-agent: prompt summary, response summary, execution time (ms), retry status
- Confidence score badge with breakdown (accuracy / completeness / clarity)
- WhatsApp-style real-time transcription — text appears word by word as you speak
- Powered by Web Speech API (no external service)
- Full session management with SQLite
- Chat history sidebar with relative timestamps
- Load any previous conversation
| Layer | Technology |
|---|---|
| Frontend | React 18, TypeScript, Tailwind CSS, Framer Motion |
| Backend | FastAPI, Python 3.10 |
| LLM | Groq API (LLaMA 3.3 70B), Gemini, Ollama |
| Database | SQLite (chat history + query cache) |
| Tunnel | NGROK |
AI_AGENT/
├── backend/
│ ├── agents/
│ │ ├── understanding_agent.py # Stage 1 — Query Decomposition
│ │ ├── research_agent.py # Stage 2 — Information Gathering
│ │ ├── analysis_agent.py # Stage 3 — Synthesis
│ │ ├── validation_agent.py # Stage 4 — Scoring + Retry Loop
│ │ └── writer_agent.py # Stage 5 — Final Output
│ ├── routers/
│ │ ├── chat.py # /api/run endpoint
│ │ ├── compare.py # /api/compare endpoint
│ │ ├── chat_sessions.py # Session CRUD
│ │ └── chat_history.py # Message History
│ ├── pipeline.py # Pipeline Orchestration + Tracing
│ ├── llm.py # Pluggable LLM (Groq / Gemini / Ollama)
│ ├── memory.py # SQLite Cache + MemoryStore
│ ├── database.py # SQLite Session / Message Helpers
│ └── main.py # FastAPI App Entry Point
└── frontend/project/
└── src/
├── App.tsx # Root: State, Routing, Compare Mode
└── components/
├── ChatMessage.tsx # Message + Confidence Badge + Trace
├── ChatInput.tsx # Input + Voice Recording
├── Sidebar.tsx # Session History
├── AgentTrace.tsx # Expandable Reasoning Panel
└── CompareView.tsx # Side-by-side LLM Comparison
---
- Python 3.10+
- Node.js 18+
- Groq API Key (free)
- NGROK account (free)
cd backend
pip install -r requirements.txt
Create .env:
GROQ_API_KEY=your_groq_api_key
GEMINI_API_KEY=your_gemini_api_key # optional, for compare mode
LLM_PROVIDER=groq
PORT=9000
python main.py
Frontend
cd frontend/project
npm install
npm run dev
Create frontend/project/.env:
VITE_API_URL=https://your-ngrok-url.ngrok-free.app
-------------------------------------------------------------------------------------------------------------------------------------------------------
📄 License
MIT