LLM WB Parser

Wildberries review analyzer powered by Claude (Anthropic). Scrapes reviews from any WB product page, classifies each one with an LLM (sentiment, topics, problems, praises, fake detection), aggregates the results, generates AI insights, and produces a ready-to-send PDF report.

How it works

WB URL
  └─► parse reviews (aiohttp, async)
        └─► LLM classification per review (Claude Haiku, concurrent)
              ├─► sentiment: positive / neutral / negative
              ├─► sentiment_score: float
              ├─► problems / praises: list of strings
              ├─► main_topics: list of strings
              └─► is_fake_likely: bool
                    └─► aggregation + clustering (scikit-learn)
                          └─► AI insights (Claude Sonnet)
                                └─► PDF report (ReportLab + matplotlib)

All LLM calls are cached in SQLite — re-running on the same product skips already-classified reviews.

Models used

Task	Model
Review classification	claude-haiku-4-5
Insights generation	claude-sonnet-4-6
Quote selection	claude-haiku-4-5

Setup

pip install -r requirements.txt
cp .env.example .env
# paste your ANTHROPIC_API_KEY into .env

Usage

python main.py --url "https://www.wildberries.ru/catalog/12345678/detail.aspx"

Options

Flag	Default	Description
--url	required	WB product page URL
--output	report.pdf	output PDF path
--max-reviews	500	max reviews to fetch (WB cap: 1000)
--client-role	селлер-конкурент	changes tone of insights: селлер-конкурент or сам владелец товара

What the report contains

• Sentiment distribution (positive / neutral / negative) with chart
• Top-10 problems and praises extracted from reviews
• Fake review percentage estimate
• Sentiment dynamics by month
• Topic distribution
• AI-generated strategic insights tailored to the client role
• Representative review quotes per topic

Project structure

├── main.py                  # CLI entrypoint
├── config.py                # models, concurrency limits, paths
├── cache.py                 # SQLite cache for LLM responses
├── parsers/
│   └── wildberries.py       # async WB scraper
├── llm/
│   ├── client.py            # Anthropic async client wrapper
│   ├── classifier.py        # per-review classification
│   ├── prompts.py           # prompt templates
│   └── quote_selector.py    # picks representative quotes
├── analytics/
│   ├── aggregator.py        # stats aggregation
│   ├── clustering.py        # topic clustering
│   └── insights.py          # AI insights generation
└── reporting/
    ├── pdf_builder.py        # assembles the PDF
    └── charts.py             # matplotlib charts

Requirements

• Python 3.11+
• Anthropic API key