docs: update README with service page links and sample library backlinks

README.md

@@ -1,77 +1,186 @@
+# E-commerce Visual Matching Dataset
+
+**Candidate match workflow dataset for product identity resolution, visual similarity, and review decision fields.**
+
+A public-safe workflow preview of how Octoparse structures AI-assisted product matching pipelines for e-commerce and pricing teams. Every row represents a candidate pair evaluation — the same structure delivered to production clients.
+
+Built by **[Octoparse Managed Data Service](https://www.octoparse.com/data-service/ai-visual-product-matching)** — managed web data pipelines for pricing intelligence and catalog teams.
+
+> **Public-safe proof, not raw client data.** Source product IDs, retailer names, and image URLs have been masked or replaced with workflow tokens. This dataset is designed to show schema, field naming, QA context, and delivery format. Final production pipelines are scoped per customer and delivered under agreed terms.
+
 ---
-pretty_name: AI Visual Matching & Product Resolution Sample Dataset
-language:
-- en
-license: other
-size_categories:
-- 1K<n<10K
-task_categories:
-- tabular-classification
-tags:
-- tabular
-- ai-visual-product-matching
-- visual-product-matching
-- product-resolution
-- ecommerce
-- automotive-aftermarket
-- candidate-retrieval
-- pre-vision-filtering
-- managed-data-service
-configs:
-- config_name: visual_matching_workflow
-  default: true
-  data_files:
-  - split: train
-    path: visual_matching_workflow_sample_1000.csv
-- config_name: product_summary
-  data_files:
-  - split: train
-    path: product_summary_sample.csv
-- config_name: data_dictionary
-  data_files:
-  - split: train
-    path: data_dictionary.csv
+
+## Why This Dataset
+
+Most product matching datasets available publicly are:
+- **Flat and unstructured** — raw scraped listings without match decisions or rejection reasons
+- **Benchmark-only** — built for model accuracy evaluation, not production workflow design
+- **Missing QA context** — no confidence scores, reject reasons, or review routing logic
+
+This dataset is different:
+- **Candidate-pair schema** — each row is a source↔candidate evaluation, not a raw product listing
+- **Match decisions included** — `match_status` captures the final output: accepted, rejected, or needs review
+- **Reject reasons structured** — `reject_reason` shows why candidates were declined, enabling workflow analysis
+- **Review notes present** — `review_note` captures QA context for borderline cases
+
 ---
-# 🖼️ AI Visual Matching & Product Resolution Sample Dataset
 
-This dataset is a sample corpus designed to train computer vision models, multimodal LLMs, and e-commerce AI agents on **Visual Matching and Product Resolution** tasks. 
+## Dataset Structure
+
+```
+ecommerce-visual-matching-dataset/
+├── visual_matching_workflow_sample.csv   ← Candidate-pair workflow rows
+├── schema.json                           ← Machine-readable field definitions
+├── data_dictionary.md                    ← Business context for every field
+└── LICENSE                              ← CC BY-NC 4.0
+```
 
-**This clean, structured sample was extracted and normalized by the [Octoparse Managed Data Service](https://www.octoparse.com/data-service/web-data-for-ai) team.**
+### Core Fields
 
-## 📊 Dataset Overview
-Training models to match identical products across different websites requires high-quality image-to-text and image-to-image pairs. Building the pipeline to extract these images, bypass anti-bot protections, and structure the metadata takes months of engineering. 
+| Field | Type | Description |
+|---|---|---|
+| `source_product_id` | string | Masked identifier for the source product being matched |
+| `candidate_product_id` | string | Masked identifier for the candidate product retrieved |
+| `similarity_score` | float | Visual similarity score (0.0–1.0) from image matching model |
+| `match_status` | string | Final decision: `accepted`, `rejected`, `needs_review` |
+| `reject_reason` | string | Structured rejection category (null for accepted matches) |
+| `review_note` | string | QA annotation for borderline or escalated candidates |
 
-We’ve provided a sample of what a production-ready visual matching pipeline looks like.
+### `match_status` Values
 
-* **Format:** JSONL / Parquet (Replace with your actual format)
-* **Domain:** E-commerce / Retail 
-* **Use Cases:** Multimodal fine-tuning, automated catalog matching, visual search training.
+| Value | Meaning |
+|---|---|
+| `accepted` | Candidate confirmed as the same or equivalent product |
+| `rejected` | Candidate declined — see `reject_reason` |
+| `needs_review` | Low confidence or conflicting signals — routed to analyst |
 
-## 🗂️ Data Structure (Schema)
-*(Note: 替换成你真实的字段)*
-* `image_url`: High-resolution source image link
-* `product_title`: Extracted product name
-* `source_platform`: Website where the data was extracted
-* `matched_id`: Unique identifier for identical products across platforms
-* `metadata`: JSON object containing variants, colors, and dimensions
+### `reject_reason` Categories
+
+| Value | Meaning |
+|---|---|
+| `different_product` | Visually similar but distinct item |
+| `different_variant` | Same product line, incompatible variant (size, color, spec) |
+| `image_quality` | Source or candidate image insufficient for visual comparison |
+| `low_confidence` | Similarity score below threshold with no confirming text signal |
+| `title_conflict` | Image match but title/model-number contradiction |
+
+---
 
-## 🚀 Need 10 Million Rows of Custom Training Data?
-Common Crawl is too noisy. Building your own scrapers is a waste of your engineering talent.
+## Quick Start
 
-If your team is building an AI agent or fine-tuning an LLM and needs highly specific, deduplicated data (text, images, or social signals from platforms like Xiaohongshu/Douyin):
+```python
+import pandas as pd
 
-**Stop building scrapers. Let us build the pipeline.**
+df = pd.read_csv("visual_matching_workflow_sample.csv")
 
-👉 **[Request a Free Custom Sample Dataset from Octoparse](https://www.octoparse.com/data-service/web-data-for-ai)**
-*We scope the project, handle the extraction, and deliver analysis-ready data to your S3/Snowflake in days.*
+# Output bucket distribution
+print(df["match_status"].value_counts())
 
-## Related Octoparse Pages
+# Rejection reason breakdown
+rejected = df[df["match_status"] == "rejected"]
+print(rejected["reject_reason"].value_counts())
+
+# Similarity score distribution by decision
+print(df.groupby("match_status")["similarity_score"].describe())
+
+# Review queue — candidates needing human judgment
+review_queue = df[df["match_status"] == "needs_review"]
+print(f"{len(review_queue)} candidates routed to review")
+
+# Accepted matches with high confidence
+high_confidence = df[(df["match_status"] == "accepted") & (df["similarity_score"] >= 0.85)]
+print(f"{len(high_confidence)} high-confidence accepted matches")
+```
+
+---
+
+## What This Dataset Answers
+
+**What is AI visual product matching?**
+Visual product matching is the process of determining whether products listed across different retailer, marketplace, or competitor sites refer to the same item — using image similarity, text signals, and structured QA to produce an explainable match decision.
+
+**Why does matching require more than image similarity?**
+A high `similarity_score` alone is not sufficient. Products can look nearly identical but differ in model number, size, or spec. This workflow combines visual similarity with title validation and structured rejection reasons to produce decisions a pricing team can act on.
+
+**What does a production matching output look like?**
+This dataset shows the candidate-pair output format: each row is a source↔candidate evaluation with a similarity score, a match decision, a rejection reason if applicable, and a review note for borderline cases. This is the data contract delivered to production clients.
+
+**How should downstream systems consume match decisions?**
+- `accepted` rows feed directly into pricing intelligence or catalog deduplication
+- `needs_review` rows route to an analyst queue for QA
+- `rejected` rows with `reject_reason` can be used to tune retrieval thresholds
+
+---
 
-This sample dataset is a public-facing preview of how Octoparse structures managed AI visual product matching workflows.
+## Use Cases
 
-- AI Visual Product Matching Service: https://octoparse.com/data-service/ai-visual-product-matching
-- Extreme Dimensions Case Study: https://octoparse.com/data-service/ai-visual-product-matching/extreme-dimensions-case-study
-- Octoparse Managed Data Service: https://octoparse.com/data-service
+### 1. Competitor Price Monitoring
+Use accepted matches to build a cross-retailer price comparison table. `source_product_id` maps your catalog; `candidate_product_id` maps competitor listings. Accepted rows give you the joined dataset for price delta analysis.
+
+### 2. Catalog Deduplication
+Use `match_status = accepted` + `similarity_score` to identify duplicate or equivalent listings in a multi-source product catalog. `reject_reason = different_variant` helps separate true duplicates from variant siblings.
+
+### 3. Matching Pipeline Evaluation
+Use the full candidate-pair schema to design or evaluate your own matching pipeline. The `reject_reason` distribution shows where retrieval methods fail and where visual scoring succeeds.
+
+### 4. Review Queue Design
+Use `needs_review` rows + `review_note` to design analyst escalation logic. The notes show what information a QA analyst needs to resolve low-confidence candidates.
+
+### 5. AI Model Training Data Structure
+Use the schema as a template for structuring your own training data. The `similarity_score` + `match_status` pairing provides a label structure for fine-tuning multimodal classifiers.
+
+---
+
+## Data Limitations
+
+| Limitation | Detail |
+|---|---|
+| Public-safe only | Source and candidate IDs, retailer names, and image URLs are masked |
+| Workflow preview | This is not a complete production export. Row count is representative, not exhaustive |
+| No raw images | Image URLs are not included in this public release |
+| Score calibration | `similarity_score` thresholds are illustrative; production thresholds are scoped per client catalog |
+| Not for benchmarking | Do not use to claim measured model accuracy on real catalog data |
+
+---
+
+## Related Resources
+
+- **Service page:** [AI Visual Product Matching Service](https://www.octoparse.com/data-service/ai-visual-product-matching)
+- **Automotive case study:** [Automotive Parts Matching — Extreme Dimensions](https://www.octoparse.com/data-service/ai-visual-product-matching/automotive-parts-case-study)
+- **Retail case study:** [Retail Product Matching Workflow](https://www.octoparse.com/data-service/ai-visual-product-matching/retail-product-matching)
+- **Data Service hub + sample library:** [Octoparse Data Service](https://www.octoparse.com/data-service#sample-library)
+- **Paired Kaggle dataset:** [ecommerce-visual-matching-dataset on Kaggle](https://www.kaggle.com/datasets/octoparsedataservice/ecommerce-visual-matching-dataset)
+
+---
+
+## Want Production-Scale Visual Matching?
+
+This is a **sample dataset**. If your team needs:
+- **Cross-retailer product matching** at scale (thousands of SKUs per day)
+- **Automotive, furniture, electronics, or apparel** catalog matching
+- **Similarity scoring + QA review routing** with structured rejection reasons
+- **Delivery to Snowflake, BigQuery, or S3** in CSV, Parquet, or JSONL
+
+→ **[Talk to Octoparse Managed Data Service](https://www.octoparse.com/data-service/ai-visual-product-matching)**
+
+We scope the matching workflow, handle retrieval and scoring, and deliver reviewed outputs your pricing or catalog team can act on directly.
+
+---
+
+## Citation
+
+```bibtex
+@dataset{octoparse_visual_matching_2025,
+  title        = {E-commerce Visual Matching Dataset},
+  author       = {{Octoparse Managed Data Service}},
+  year         = {2025},
+  publisher    = {Hugging Face},
+  url          = {https://huggingface.co/datasets/Octoparse/ecommerce-visual-matching-dataset},
+  license      = {CC BY-NC 4.0},
+  note         = {Public-safe candidate-pair workflow dataset for AI visual product matching and pricing intelligence}
+}
+```
+
+---
 
-Need a scoped AI visual matching POC?  
-Contact Octoparse Managed Data Service to discuss your target catalog, source websites, candidate retrieval requirements, and output format.
+*Built by [Octoparse Managed Data Service](https://www.octoparse.com/data-service) · [LinkedIn](https://www.linkedin.com/company/octoparse)*