Data Chat (NL2SQL)

An AI-powered natural language interface for querying your Django database directly from the admin.

Overview

Data Chat allows staff users to ask questions in plain English and receive SQL-backed answers. The system:

1. Converts natural language to SQL using an LLM
2. Validates queries against a security policy (read-only, allowed tables)
3. Executes the query and returns results
4. Summarizes results in natural language

Quick Setup

1. Add to INSTALLED_APPS

# settings.py
INSTALLED_APPS = [
    ...,
    "automate_datachat",
]

2. Register Models

In your app's apps.py:

from django.apps import AppConfig

class MyAppConfig(AppConfig):
    name = 'myapp'

    def ready(self):
        from automate_datachat.registry import DataChatRegistry
        from .models import Product, Order, Customer

        DataChatRegistry.register(
            Product,
            include_fields=["id", "name", "price", "category", "created_at"],
            tags=["inventory"]
        )

        DataChatRegistry.register(
            Order,
            include_fields=["id", "customer_id", "total", "status", "created_at"],
            tags=["sales"]
        )

3. Run Migrations

python manage.py migrate automate_datachat

Admin Widget

The chat widget appears automatically in Django Admin for staff users (bottom-right corner).

Features

• Conversational: Maintains context across questions
• SQL Display: Shows generated SQL for transparency
• History: Persists messages across sessions

Prompt Management

Prompts are stored in the database and editable via Admin.

Default Prompts

• datachat_sql_generator - Converts natural language to SQL
• datachat_summarizer - Summarizes query results

Template Variables

Prompts use Jinja2 templating:

Variable	Description
{{ schema }}	Database schema (tables, columns)
{{ history }}	Conversation history
{{ question }}	User's current question
{{ results }}	Query results (for summarizer)

LLM Logging

All LLM interactions are logged to LLMRequest:

Field	Description
input_payload	Full messages sent to LLM
output_content	Raw LLM response
input_tokens	Tokens consumed
latency_ms	Response time
cost_usd	Cost (if provider returns it)

View logs in Admin → Automate LLM → LLM Requests.

Security

Read-Only Queries

Only SELECT statements are allowed. INSERT, UPDATE, DELETE are blocked.

Table Whitelisting

Only registered models are exposed. Unregistered tables cannot be queried.

SQL Validation

All generated SQL is validated before execution using SQLPolicy.

API Endpoints

Endpoint	Method	Description
/datachat/api/chat/	POST	Send a question
/datachat/api/history/	GET	Get message history

Configuration

# settings.py
DATACHAT_CONFIG = {
    "max_results": 1000,  # Max rows returned
    "default_provider": "openai",  # LLM provider to use
}