Nabil Asofi

Library / crawl4ai

Review-Mining Pipeline

Pull every review you have, cluster the complaints, and rank them by how much revenue they're costing.

Setup time
30 min
Stack
crawl4aiLLM APIGoogle Sheets
Replaces
Reading reviews until you feel like you understand, then acting on the loudest one

What it does, in order

Review-Mining Pipeline — pipeline of 5 steps Pipeline of 5 steps: 1. Crawl the review pages (crawl4ai); 2. Extract the claims (LLM API); 3. Cluster into themes (LLM API); 4. Rank frequency and severity (LLM API); 5. Write the findings (Google Sheets). 01 Crawl the review pages crawl4ai 02 Extract the claims LLM API 03 Cluster into themes LLM API 04 Rank frequency and severity LLM API 05 Write the findings Google Sheets Review-Mining Pipeline — pipeline of 5 steps Pipeline of 5 steps: 1. Crawl the review pages (crawl4ai); 2. Extract the claims (LLM API); 3. Cluster into themes (LLM API); 4. Rank frequency and severity (LLM API); 5. Write the findings (Google Sheets). 01 Crawl the review pages crawl4ai 02 Extract the claims LLM API 03 Cluster into themes LLM API 04 Rank frequency and severity LLM API 05 Write the findings Google Sheets
5 steps. The final node is the artifact you end up with.

What this does

Reviews get collected from the platforms that matter for your category — your own store, plus whichever of G2, Trustpilot, Amazon or the app stores your buyers actually use. They’re cleaned, clustered by the underlying complaint rather than the words used, and ranked by frequency and sentiment weight.

The output is a ranked list. Each entry has the cluster name, how many reviews support it, and three verbatim quotes you can put in front of someone who needs convincing.

The distinction that matters

“The shipping was slow” and “it arrived late and I’d already left for holiday” are the same cluster. “The app logged me out” and “I lost my cart again” are usually the same cluster. Word frequency misses this; it counts the phrasing, not the problem.

What you get

  • The workflow file with the source adapters
  • The clustering prompt set, including the instruction that keeps it from producing twelve clusters that are all “usability”
  • A setup document: source configuration, rate-limit handling, and the verbatim-selection rules
  • A note on where this pays for itself — usually the review theme that never made it into a support ticket because the customer just left

Honest limits

It reads text. A complaint that lives only in a support call, a returns reason code, or a one-star rating with no words attached needs a different input. The pipeline flags those as silent, but it cannot mine them.