Problem
A food-delivery merchant on Grab can have hundreds or thousands of reviews. Consumers read the first few and get an incomplete picture, and the star average does not say why two restaurants with the same rating differ in food quality, delivery, portion size or value. Merchants face the same wall when they try to find recurring complaints by hand. A wrong or one-sided summary changes how consumers see a merchant.
Example from the article: the phrases "too cold", "not warm enough" and "lukewarm" all describe one food-temperature issue, which keyword search may treat as separate queries.
Solution in brief
Grab samples reviews so that the sample follows the merchant's rating distribution, asks an LLM for an overall summary and per-topic summaries (taste, packaging, portion size), and has a second LLM check each summary against the sampled reviews before publication. A monitoring job regenerates the summary when enough new reviews or enough negative reader feedback accumulate.
Questions this case answers
- How to generate a faithful short summary from a large, growing pile of user-written texts when the model can only read a sample of them
- How to keep a generated summary up to date when new texts and negative reader feedback keep arriving
- How to turn hundreds of written reviews of a merchant into a short summary that readers can trust?
- How to pick a review sample for an LLM so that one rating group does not dominate the summary?
Tools and methods
- Stratified sampling by rating — pick a review subset that keeps the merchant's rating distribution
- Sentiment model — scores each written review from 1 to 5 so reviews can be stratified
- ChatGPT — writes the overall and topic-based summaries (per Fig. 2)
- Gemini — separate validator: checks summaries and analyses reader feedback (per Fig. 2 and 3)
- Threshold triggers — start regeneration on new-review count or negative-feedback count
Google or a one-time email code. No password.