Apparel Product Review Issue & Trend Analyzer

Turn authorized English apparel reviews into six-topic issue counts with exact source quotes, per-product shares, and cautious time-window comparisons. This public Actor analyzes supplied text; it does not scrape websites or discover customers. It charges $0.03 for each successfully completed review-analysis report. Store availability and technical tests do not prove paid demand or profitability.

Features and scope

Semantic mode uses glm-5.3-flash through Tencent Cloud TokenHub's Guangzhou pay-as-you-go endpoint. Topics: sizing too small/large, disliked material, actual damage or care-related deterioration, delivery problems, and factual description mismatches. Price, style, and unflattering shape are outside this taxonomy. Rating alone never establishes a complaint.

Findings include product, topic, mentions, shareOfReviews, lowRatingMentions, periodReviewCounts, periodMentions, periodRates, rateChangePercentagePoints, trend, and up to three exact evidence objects. The complete JSON also contains summary, engine, version, asOf, limitations, and modelAccounting. REPORT.csv provides a spreadsheet-safe summary.

Quick start

1. Supply authorized English review objects with text or body, not both. Optional fields: id, product, rating (1-5), and ISO date. Remove personal information first.
2. The default rules mode is a key-free smoke test using synthetic example input. For AI analysis, choose semantic and provide your Guangzhou TokenHub pay-as-you-go key in the encrypted tokenhubApiKey field. Never put keys inside reviews. Rules mode is a limited offline diagnostic, not equivalent AI analysis.
3. Set windowDays from 1 to 365 (default 30); optionally set asOf for reproducibility.
4. Read the dataset, complete OUTPUT JSON, or REPORT.csv.

Example, no key needed:

{"mode":"rules","reviews":[
  {"id":"demo-1","product":"Demo shirt","text":"Too small across the shoulders.","rating":2},
  {"id":"demo-2","product":"Demo shirt","text":"Perfect fit and comfortable fabric.","rating":5}
]}

Limits, costs, and failures

Maximum 500 submitted records, 100 unique reviews, and 30,000 unique text characters per run; each review text is limited to 6,000 characters. Identical normalized text is deduplicated within each product. Unknown input fields and caller-supplied URLs or endpoints are rejected. No browser, shell, remote tools, or arbitrary URL fetching are used.

Semantic mode batches 20 reviews, caps output at 4,000 tokens per request, and preflights the run against an estimated CNY 0.20 model budget. This estimate is not a provider-enforced hard cap. Requests are sequential, with no automatic model retries or silent rules fallback. A network failure, incomplete output, unsupported label, or ungrounded quote fails the run; requests already sent can still cost money.

Customers bring their own key and pay Tencent model fees directly. The public Store price is one $0.03 review-analysis event per successfully completed report. Apify platform usage, storage, transfer, and external model fees are separate according to the customer's accounts and plans. Before any model request, the Actor requires PAY_PER_EVENT pricing with review-analysis at no more than $0.05. It requests exactly one developer charge only after OUTPUT and REPORT.csv exist, records billing intent before charging, uses the Apify run ID for idempotency, and withholds dataset publication if charging is unconfirmed. An unresolved pending intent blocks retry charging until manual reconciliation. Invalid-input tests completed with no developer charge. Reported model cost is an estimate, not a verified invoice or total bill.

Public runs require BYOK. Creator environment-key fallback is blocked whenever the Actor is public or the runner is not its owner; no creator key is required for public operation.

Quality and privacy

A diagnostic evaluation of 60 public historical reviews produced 55 exact matches, 95.1% micro precision, and 92.9% recall after case-only evidence alignment. This was not independent human validation, a production guarantee, or a complaint-prevalence estimate.

Common emails, phone patterns, and sk-style keys are redacted before model calls and reports, but this is not comprehensive PII removal. Raw input remains subject to Apify storage behavior; sanitized text is sent to Tencent Cloud. Submit only lawful, authorized text without unnecessary sensitive information. Exact evidence proves grounding, not classification correctness. Do not profile individual reviewers.

Trends require ten dated reviews in each comparison window and four issue mentions. Missing, invalid, or future dates cannot establish trends. Shares use all unique product reviews. These are descriptive signals, not significance tests or causal conclusions.

Troubleshooting and development

For failures, check Guangzhou key access, model availability, balance or quota, restart settings, and the safe error record. No alternate model is selected silently. On interruption, the handler attempts to write ABORT_STATUS and prevent further batches, but a direct platform abort can terminate the container first and an in-flight request can still cost money. No background scheduler, outreach, customer acquisition, or payout automation is configured.

Core offline tests: Python 3.11+, python3 -m unittest discover -v. Runtime dependencies are in requirements.txt; Docker entry point is main.py. Standalone analyzer.py is only the rules diagnostic. Verify the current run estimate and Store pricing before starting.