LLM-based E-commerce Intelligent Analytics Assistant: Fusion Practice of RAG and NL-to-SQL

This article introduces an e-commerce data analysis system that combines Large Language Model (LLM), Retrieval-Augmented Generation (RAG), and Natural Language to SQL (NL-to-SQL) technologies, aiming to enable business personnel to obtain in-depth data insights through natural language. The project is maintained by Rihanashariff and was published on the GitHub platform on June 6, 2026, with the original title "LLM-Powered-Analytics-Assistant-with-RAG."

In e-commerce operations, business personnel face two major challenges: First, massive structured data (transactions, user behavior) requires complex SQL skills to extract value, and relying on the data team leads to long waiting cycles; second, unstructured data such as customer reviews is difficult to perform semantic understanding and sentiment analysis through traditional BI tools. The need is to build a natural language interaction system that allows business personnel to directly ask questions to the data and get visual results.

Data Layer

Uses the public dataset of Brazilian e-commerce Olist (about 100,000 orders, 8 tables, 40,000 customer reviews, 2016-2018), stored in SQLite.

Core Engine

• NL-to-SQL module: Converts English questions to SQL
• RAG retrieval module: Processes unstructured data based on FAISS vector database
• Hybrid reasoning module: Intelligently coordinates SQL and RAG
• Visualization module: Automatically generates Plotly charts

Technology Stack

LLM API (Groq), embedding model (Sentence Transformers), vector database (FAISS), frontend (Streamlit), data processing (Pandas/NumPy)

Structured Data Query (SQL Mode)

For metrics like order volume and sales, automatically generates SQL, e.g.:

• Top 5 product categories with the highest revenue
• Average order delivery time

Unstructured Data Analysis (RAG Mode)

Processes customer reviews through FAISS semantic retrieval + LLM summary/sentiment analysis, e.g.:

• Customer reviews on delivery services
• Main reasons for low-rated reviews

Hybrid Analysis Mode (HYBRID Mode)

Combines structured and unstructured data, e.g.:

• Impact of delivery time on customer satisfaction
• Product categories with low sales and poor reviews

Hybrid Analysis Mode (HYBRID Mode)

Combines structured and unstructured data, e.g.:

• Impact of delivery time on customer satisfaction
• Product categories with low sales and poor reviews

Intelligent Routing

LLM automatically judges the query type:

• Only needs structured data → SQL mode
• Only needs text analysis → RAG mode
• Combination of both → HYBRID mode

Visualization

Automatically recommends chart types (bar charts, line charts, etc.), generates Plotly interactive charts, and achieves the experience of "get charts immediately after asking questions."

Project Structure

• data/: Database and vector indexes
• sql/: NL-to-SQL conversion
• rag/: Embedding generation and semantic retrieval
• llm/: Routing, sentiment analysis, result synthesis, chart generation
• app.py: Streamlit interface

Practical Value

1. Natural language interface lowers the threshold for data analysis
2. RAG effectively handles unstructured text
3. Hybrid reasoning answers complex business questions
4. Open-source tech stack (Groq replaces OpenAI, FAISS replaces commercial vector databases) controls costs

This project provides a practical reference implementation for e-commerce data analysis, showing the complete path from data preparation to interface design. With the improvement of LLM capabilities and cost reduction, conversational data analysis is expected to become a standard configuration of enterprise BI tools, and its open-source implementation provides valuable practical experience for the industry.