SCSI-SLM: Consumer Voice to Engineering Insight

Official implementation for the paper:

Mapping Consumer Voice into Engineering Insight: A Structured Language Model-Driven Design Support Framework for Electric Vehicles

Published in Journal of Engineering Design
Article page: https://www.tandfonline.com/doi/full/10.1080/09544828.2026.2639933
DOI: https://doi.org/10.1080/09544828.2026.2639933

Abstract

This repository contains the code, processed artifacts, and demo application for the SCSI-SLM framework. The project studies how large language models and structured retrieval can transform unstructured consumer reviews into engineering design insight for electric vehicles. The full pipeline includes:

1. structured semantic encoding of user reviews,
2. product-side and user-side modeling,
3. engineering knowledge graph construction,
4. a hybrid RAG application for interactive design support.

Highlights

• End-to-end research pipeline from raw EV review data to engineering insight generation.
• Modular implementation aligned with the paper sections.
• Included processed outputs, analysis reports, figures, and a runnable RAG demo.
• Support for both vector retrieval and graph retrieval in the final application.

Framework Overview

Repository Structure

Path	Paper Section	Purpose
00_Raw_Data/	Sec. 3.1.1	Raw consumer review datasets used in the study.
01_SSE_Analysis/	Sec. 3.1	Data cleaning, LLM tag extraction, and engineering-dimension mapping.
02_User_Modeling/	Sec. 3.2	Importance-performance analysis and user preference clustering.
03_Knowledge_Graph/	Sec. 3.3	Neo4j-based engineering design knowledge graph construction.
04_RAG_APP/	Sec. 4.3	Interactive hybrid retrieval and reasoning system.
docs/images/	Figures	Representative figures used in the manuscript and README.

Requirements

• Python 3.8+
• Docker and Docker Compose for Neo4j deployment
• OpenAI-compatible API key for LLM-powered steps
• Neo4j database for graph construction and graph retrieval

Install the Python dependencies from the project root:

pip install -r requirements.txt

Core dependencies include:

• pandas, numpy, scikit-learn
• jieba
• matplotlib, seaborn, plotly
• langchain, langchain-openai, langchain-community, openai
• chromadb, neo4j, streamlit

Data and Artifacts

The repository includes both source data and intermediate outputs from the research workflow.

• Raw review data is stored in 00_Raw_Data/.
• Cleaned review outputs are stored in 01_SSE_Analysis/1_Data_Preprocessing/outputs/.
• User modeling outputs are stored under 02_User_Modeling/.../outputs/.
• The knowledge graph module contains graph-building scripts and Neo4j setup files.

If you use the data or code in academic work, please cite the paper listed below.

Reproducing the Pipeline

The project is organized as a staged workflow rather than a single training script.

1. Structured Semantic Encoding

This stage converts raw consumer reviews into structured engineering tokens.

Key scripts:

• 01_SSE_Analysis/1_Data_Preprocessing/cleaning_pipeline.py
• 01_SSE_Analysis/2_Dimension_Construction/tag_extraction_refinement.py
• 01_SSE_Analysis/2_Dimension_Construction/dimension_mapping.json

Expected outputs:

• cleaned comments
• refined feature tags
• mapped engineering dimensions

2. Product and User Modeling

This stage derives both product-side and user-side signals from the structured review data.

Product-side analysis:

• 02_User_Modeling/Product_IPA_Analysis/ipa_quantification.py
• outputs include car_model_scores.csv, feature_statistics.json, and per-model IPA figures

User-side analysis:

• 02_User_Modeling/User_Preference_Clustering/preference_profiling.py
• 02_User_Modeling/User_Preference_Clustering/persona_visualization.py
• outputs include user vectors, cluster characteristics, reports, and visualization figures

3. Knowledge Graph Construction

This stage organizes the extracted entities and relations into an engineering design knowledge graph.

Key files:

• 03_Knowledge_Graph/main.py
• 03_Knowledge_Graph/scripts/init_database.cypher
• 03_Knowledge_Graph/src/knowledge_graph_builder.py
• 03_Knowledge_Graph/docker-compose.yml

4. Hybrid RAG Application

The final stage exposes the research outputs through an interactive interface.

Key files:

• 04_RAG_APP/app.py
• 04_RAG_APP/run.py
• 04_RAG_APP/load_vector_data.py
• 04_RAG_APP/core/rag_engine.py

Demo

To run the interactive application locally:

1. Configure environment variables

Create a root-level .env file based on .env.example.

Required variables include:

• OPENAI_API_KEY
• NEO4J_URI
• NEO4J_USERNAME
• NEO4J_PASSWORD
• NEO4J_DATABASE

2. Start Neo4j

From 04_RAG_APP/ or 03_Knowledge_Graph/, start the database with Docker Compose:

docker-compose up -d neo4j

3. Initialize vector data if needed

cd 04_RAG_APP
python load_vector_data.py

4. Launch the app

cd 04_RAG_APP
streamlit run app.py

Or:

cd 04_RAG_APP
python run.py

The default web interface is available at http://localhost:8501.

Representative Outputs

Importance-Performance Analysis

User Preference Clustering

Hybrid RAG Interface

Application Scenario

The SCSI-SLM framework supports multiple applications in electric vehicle product development.

Typical scenarios include:

• Product manager decision support based on consumer feedback
• Engineering design insight mining from large-scale reviews
• Knowledge-grounded product planning

Related Projects

An example application prototype built on this framework:

EV Product Manager Decision Support System (EV-PM-DSS)

https://github.com/DonkeyKing01/EV-PM-DSS

Reproducibility Notes

• This repository provides the full research pipeline as modular code and includes many intermediate outputs.
• Some stages rely on external LLM services and therefore require valid API credentials.
• The final demo depends on artifacts generated by the earlier modules and on a running Neo4j instance.
• Exact outputs may vary across model providers, prompts, or data versions.

Citation

If you find this repository useful, please cite:

@article{Jin2026Mapping,
  title   = {Mapping Consumer Voice into Engineering Insight: A Structured Language Model-Driven Design Support Framework for Electric Vehicles},
  author  = {Qingyang Jin and Luyao Wang and Wenyu Yuan and Danni Chang},
  journal = {Journal of Engineering Design},
  year    = {2026},
  doi     = {10.1080/09544828.2026.2639933}
}