Medical QueryBot - Task 1

A Natural Language to SQL to Answer system for medical data queries.

Features

• Natural language question processing
• SQL query generation using LLM
• SQLite database integration
• Streamlit web interface
• Safe SQL validation (SELECT-only, single statement)
• ETL to ingest JSON/CSV into SQLite
• JSON response with metadata
• Minimal/Hybrid RAG (keyword + optional embeddings)
• Retry/repair on SQL validation failure
• Execution time metrics and detailed metadata

Setup

1. Dependencies

pip install -r requirements.txt

2. Environment Create .env from the example and fill values:

copy .env.example .env   # Windows
# or
cp .env.example .env     # macOS/Linux

Edit .env:

OPENAI_API_KEY=sk-...
DB_PATH=E:\JSL_TASK_1\app\db\demo.sqlite

3. Ingest data (one-time)

python run_etl.py

4. Run UI

streamlit run app/ui/streamlit_app.py

Usage

Ask medical questions in natural language and get SQL-generated answers.

Example questions:

• Kaç hasta var?
• Cinsiyete göre hasta sayıları nedir?
• Son 7 günde kaç yatış yapılmış?

Response shape:

{ "sql": "...", "answer": "...", "meta": { "results": {"row_count": 12, "columns": ["col1", "col2"]}, "validation": {"is_valid": true}, "database": {"tables_used": ["JSON_PATIENTS"], "query_type": "COUNT_QUERY", "complexity": "SIMPLE"} } }

Docker (optional)

Build and run:

docker build -t medical-querybot .
docker run --env-file .env -p 8501:8501 medical-querybot

Task 2 Enhancements (kept simple)

• Minimal RAG (keyword + TR→EN mapping) for schema context
• Optional Hybrid RAG (embeddings + keyword) via environment flags
• Retry/repair: if SQL validation fails, one guided regeneration
• Observability: execution time (ms), validation status, tables used
• UI shows: Relevant Schema, Exec Time (ms), Retry banner

Environment flags

Add to .env as needed:

# Hybrid RAG (optional)
RAG_HYBRID=1                 # 1: use embeddings+keyword, 0: keyword-only
RAG_TOP_K=3                  # how many schema snippets to include
RAG_ALPHA=0.7                # semantic score weight (0..1)
EMBEDDING_MODEL=text-embedding-3-small

# LLM controls (optional)
OPENAI_MODEL=gpt-3.5-turbo
OPENAI_TEMPERATURE=0.1

Tests

Quick scripts to validate behavior:

# Relevant schema (RAG) and full pipeline check
python test_rag_schema.py

# Retry/repair flow (should show Retried: True for at least one case)
python test_retry_flow.py

# Hybrid vs keyword-only retrieval (requires OPENAI_API_KEY)
python test_hybrid_rag.py

Notes

• Postgres + vector store are optional for Task 2; SQLite + minimal Hybrid RAG is sufficient.
• If Hybrid RAG is disabled (RAG_HYBRID=0), keyword-only retrieval is used as fallback.