Ecommerce ops · Production

From reviews to insights: Building analytic applications with Large Language Models in e-commerce

The problem

Conventional ML workflows for review analytics require collecting labeled datasets and training separate models for each task, with limited explainability and significant time investment.

First attempt

The multi-model conventional approach was intensive, time-consuming, and provided limited explainability, making it difficult for stakeholders to understand how sentiment decisions were reached.

Workflow diagram · grounded in source
1
Reviews API ingestion
integration
“beginning with extracting reviews using a reviews API”
2
LLM sentiment and aspect extraction
ai_action
“The meticulously crafted prompt in the code below instructs the model on multiple tasks while ensuring a properly formatted output for easier downstream processing. This includes handling reviews with no expressed sentiment, lengthy off-…”
3
Few-shot topic clustering
ai_action
“The approach above employs a few-shot prompt template, providing the model with explicit instructions and a handful of examples in the desired input-output format”
Reported outcome

Using LLMs enables a single model to handle multiple tasks with significant speed improvements, and provides stakeholders with greater transparency by highlighting specific review sections contributing to each sentiment assignment.

Reported metrics
Application development speedsignificant speed improvements
Fine-tuning time and computesaving time and computational resources
Reported stack
OpenAIreviews API
Source
https://medium.com/data-science-at-microsoft/from-reviews-to-insights-building-analytic-applications-with-large-language-models-in-e-commerce-ad28ee60e2a7
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Using LLMs enables a single model to handle multiple tasks with significant speed improvements, and provides stakeholders with greater transparency by highlighting specific review sections contributing to each sentime…

What tools did this team use?

OpenAI, reviews API.

What results were reported?

Application development speed: significant speed improvements; Fine-tuning time and compute: saving time and computational resources (source-reported, not independently verified).

What failed first in this deployment?

The multi-model conventional approach was intensive, time-consuming, and provided limited explainability, making it difficult for stakeholders to understand how sentiment decisions were reached.

How is this ecommerce ops AI workflow structured?

Reviews API ingestion → LLM sentiment and aspect extraction → Few-shot topic clustering.