Designed & Developed By

Nisrin Dhoondia

Retail Insights Assistant, a LLM Multi-Agent System (Blend360 GenAI Assignment)

This project is a GenAI-powered Retail Insights Assistant that:

• Cleans large inconsistent retail CSV datasets
• Loads them into DuckDB
• Also creates master sales table
• Supports Summaries and Conversational Q&A
• Uses Local Mistral (Ollama) for SQL generation
• Provides a Streamlit UI

The repository contains a fully working Retail Insights Assistant built using:

• Local LLMs (Mistral via Ollama)
• Streamlit UI
• DuckDB as the analytics engine
• Multi-Agent architecture:
  • LanguageToSQLAgent (LLM generates SQL)
  • SQLExecutionAgent
  • ValidationAgent

Features

• Auto-renames messy filenames

• Standardizes columns (lowercase, snake_case)

• Cleans values

• Loads all cleaned CSVs into DuckDB automatically

• Two unified structures created:

  – Each CSV table

  – master_sales table (merging compatible sales tables)

• Natural language to SQL using local Mistral model

• Fuzzy matching for dirty retail data

• Streamlit chatbot UI

  Modes:

  • Summarization Mode simple table summaries
  • Conversational Q&A Mode LLM-generated SQL with result viewer

• DuckDB execution engine

• Multi-Agent Architecture

  Agents:

  • LanguageToSQLAgent converts natural language to SQL
  • ValidationAgent ensures only safe SELECT queries
  • SQLExecutionAgent executes SQL on DuckDB

Folder Structure

Blend360_Retail_Insights/

│

├── app.py

├── utils.py

├── agents.py

├── requirements.txt

├── README.md

│

├── data/

│ ├── (All original CSV files given by HR)

│

├── presentation/

│ ├── Architecture_Pipeline_PDF.pdf

│ ├── Architecture_Pipeline_PPT.pptx

│ ├── Streamlit_Demo.mp4

│ ├── Sample_Questions.md

│

└── .gitignore

Running the Project

1. Install dependencies

pip install -r requirements.txt

2. Install & run Mistral (Ollama)

ollama run mistral

3. Start the Streamlit App

streamlit run app.py

How to Test

Step 1 - Load Data

• Click on the sidebar controls

Clean & (Re)Load all CSVs into DuckDB

Create master_sales (merge compatible sales tables)

Step 2 - Summarization

• Select a table

• Click Generate Summary

• The LLM generates:

  – Column-level insights

  – Patterns and trends

  – Anomalies

  – Business interpretations

  – Data quality issues

• Output:

  A structured, human-readable dataset summary

Step 3 - Conversational Q&A

• Examples you can type:

  1. What is the total amount of Amazon sales report?

  Table: All (Select All from dropdown list)

  2. Which are the states the product shipped?

  Table: amazon_sale_report (Select amazon_sale_report from dropdown list)

  3. How many null values in this table?

  Table: cloud_warehouse_compersion_chart (Select cloud_warehouse_compersion_chart from dropdown list)

  4. Which are the Increff in each Shiprocket?

  Table: cloud_warehouse_compersion_chart (Select cloud_warehouse_compersion_chart from dropdown list)

  5. What are the different expenses and its count?

  Table: expense_iigf (Select expense_iigf from dropdown list)

  6. What is the total gross amount for each customer?

  Table: international_sale_report (Select international_sale_report from dropdown list)

  7. What is the total quantity for each SKU?

  Table: master_sales (Select master_sales from dropdown list)

  8. What are the different categories and catalogs available?

  Table: may2022 (Select may2022 from dropdown list)

  9. Each catalog has how many categories?

  Table: may2022 (Select may2022 from dropdown list)

  10. What are the different catalogs available?

  Table: p__l_march_2021 (Select p__l_march_2021 from dropdown list)

  11. What are the different categories available?

  Table: sale_report (Select sale_report from dropdown list)

  12. What are the different sources?

  Table: unified_sales (Select unified_sales from dropdown list)

• The LLM will:

  • Generate SQL
  • Validate SQL
  • Execute SQL
  • Display results as data table & JSON

Limitations and Future Improvements

Limitations

• Fuzzy SQL rules may require tuning
• LLM runs locally (Mistral via Ollama)
• Deep summaries are slow on large data

Possible Improvements

• Add important column auto-detection for faster summaries
• Use metadata vector DB for column-search optimization