📊 BI Dashboard AI — Conversational Business Intelligence

Ask any question about your data in plain English. Get instant, interactive charts — no SQL, no coding required.

---

🎯 What It Does

BI Dashboard AI lets non-technical users — executives, analysts, anyone — upload a CSV file and immediately start asking business questions in plain English. The system:

1. Understands your question using LLaMA 3.3 70B via Groq
2. Generates accurate SQL against your actual data
3. Validates the SQL before running it (no hallucinated columns)
4. Visualizes results as interactive Plotly charts
5. Summarizes findings in plain English — no jargon

---

✨ Features

Feature	Description
💬 Conversational UI	Chat-style interface with message history and follow-up questions
📁 CSV Upload	Upload any CSV — auto-detects encoding and delimiter
📊 Auto Chart Selection	Picks bar, line, pie, scatter, histogram based on data shape
🔒 Hallucination Prevention	3-layer validation — prompt injection, SQLite EXPLAIN, empty result detection
📝 Plain-English Summaries	Non-technical summaries with real statistics from actual data
🔢 Single Value Cards	Clean stat cards for single-value answers instead of broken charts
❓ Vague Query Handling	Asks for clarification with suggested questions instead of guessing
⬇️ Data Export	Download any chart's data as CSV

---

🏗️ Architecture

User (Natural Language)
         │
         ▼
┌─────────────────────┐
│   Streamlit UI      │  ← Port 8501  (frontend_template.py)
│   Chat interface    │
│   Plotly charts     │
└──────────┬──────────┘
           │ HTTP POST /query
           ▼
┌─────────────────────┐
│   FastAPI Backend   │  ← Port 8000  (backend_template.py)
└────┬────────────────┘
     │
     ├──► Groq API (LLaMA 3.3 70B)
     │         └── Generates SQL + chart spec
     │
     ├──► SQLite EXPLAIN validation
     │         └── Rejects hallucinated column names
     │
     └──► SQLite Query Execution
               └── Returns real data → Plotly → User

Pipeline: Text → Groq/LLaMA → SQL → Validate → SQLite → Plotly → Streamlit

---

🚀 Quick Start

Prerequisites

• Python 3.9+
• A Groq API key (free at console.groq.com)

Run in one command

python run.py

That's it. The script will:

1. Install all dependencies automatically
2. Copy backend and frontend files
3. Start FastAPI on port 8000
4. Start Streamlit on port 8501
5. Open your browser to http://localhost:8501

Files required (keep all three together)

your-folder/
├── run.py                  ← Entry point
├── backend_template.py     ← FastAPI backend
└── frontend_template.py    ← Streamlit frontend

---

📁 Project Structure

├── run.py                  # Single-command launcher
├── backend_template.py     # FastAPI app — /query, /upload-csv, /schema
├── frontend_template.py    # Streamlit chat UI with Plotly charts
├── backend/
│   └── main.py             # Auto-generated from backend_template.py
├── frontend/
│   └── app.py              # Auto-generated from frontend_template.py
└── data/
    └── *.db                # Uploaded CSV stored as SQLite

---

🛠️ Tech Stack

Layer	Technology
LLM	Groq — LLaMA 3.3 70B Versatile
Backend	FastAPI + Uvicorn
Database	SQLite (auto-created from uploaded CSV)
Frontend	Streamlit
Charts	Plotly Express + Graph Objects
Data Processing	Pandas + NumPy

---

💬 Example Queries

Simple

"What is the average price of all vehicles?"

Intermediate

"Show me prices of 3 Series vehicles in 2018 broken down by fuel type"

Complex

"Show me the trend of average mileage and price over the years for diesel vehicles"

Follow-up (Conversational)

"Now filter this to only show vehicles above 30 MPG"

Hallucination test

"Show me total sales revenue by region" → blocked with explanation

Vague query

"Show me something cool" → asks for clarification with suggestions

---

🔒 Hallucination Prevention

The system uses 3 hard gates to prevent the AI from making up data:

1. Prompt injection — exact column names are injected into every prompt so the LLM sees the real schema
2. SQLite EXPLAIN validation — every generated SQL is validated by SQLite itself before execution; invalid column/table names are caught and blocked
3. Empty result detection — if a valid SQL returns zero rows, the user is informed rather than shown an empty chart

---

📡 API Endpoints

Method	Endpoint	Description
GET	/health	Health check
GET	/schema	View current database schema
POST	/upload-csv	Upload CSV → SQLite
POST	/query	Natural language → charts
POST	/inspect-file	Debug file parsing
GET	/sample-queries	Get example questions

POST /query

{
  "prompt": "Show me average price by model",
  "conversation_history": [],
  "use_uploaded_db": true
}

---

⚙️ Configuration

Update the Groq API key in backend_template.py:

GROQ_API_KEY = "your-groq-api-key-here"

---

📊 Evaluation Criteria Coverage

Criterion	Implementation
Accuracy — Data Retrieval (40%)	Schema-aware SQL with unique value injection and TRIM/LIKE for text matching
Accuracy — Chart Selection	LLM instructed with chart type rules; line for time-series, pie for part-of-whole, etc.
Accuracy — Error Handling	3-layer hallucination prevention; vague queries trigger clarification
Aesthetics — Design (30%)	Clean light theme, gradient header, card-based layout
Aesthetics — Interactivity	Plotly tooltips, zoom, pan, legend toggle on every chart
Aesthetics — User Flow	Loading spinner, chat-style messages, instant feedback
Architecture (30%)	Full Text→LLM→SQL→SQLite→Plotly pipeline
Prompt Engineering	Schema injection, unique value listing, TRIM/LIKE rules, temperature=0
Hallucination Handling	Explicitly reports when data is unavailable; never fabricates
Bonus — Follow-ups (+10%)	Full conversation history passed to LLM each turn
Bonus — CSV Upload (+20%)	Upload any CSV with auto encoding/delimiter detection

---

🙏 Built With

• Groq — blazing fast LLM inference
• Meta LLaMA 3.3 — open source LLM
• Streamlit — rapid frontend
• FastAPI — modern Python API
• Plotly — interactive charts

---

Built for the Qualifier — Conversational AI for Instant Business Intelligence Dashboards