Retail Analytics Copilot — DSPy + LangGraph

Graph Design

• The agent graph includes: Router → Retriever → Planner → NL2SQL → Executor → Synthesizer → Repair Loop → Finish.
• Router selects rag, sql, or hybrid based on question keywords.
• Planner extracts date ranges, KPI formulas, and product categories from retrieved doc chunks.
• The repair loop retries SQL generation up to 2 times if execution fails or output is malformed.

DSPy Optimization

• Optimized Module: NL→SQL (Chain-of-Thought + BootstrapFewShot)
• Metric: SQL validity (SQL parses + executes successfully)
• Before optimization: 5% (1 out of 20 examples)
• After optimization: 85% (17 out of 20 examples)

This optimization improved SQL generation quality and reduced repair-loop activations.

Assumptions & Trade-offs

• CostOfGoods approximation: Because Northwind has no cost data, Gross Margin uses
  CostOfGoods = 0.7 × UnitPrice.
• Date parsing: All date ranges (e.g., “Summer Beverages 1997”) come from RAG documents only.
• Category detection: Categories inferred by keyword matching in question + docs.