Automate Product Tagging with AI on GPU

What You’ll Build

In under an hour, you will have a product tagging pipeline that analyses product images and descriptions, assigns category hierarchies, extracts attributes like colour, material, and size, generates SEO-friendly search tags, and maps products to your custom taxonomy. Processing 10,000 SKUs takes under 30 minutes on a dedicated GPU server, turning a weeks-long merchandising project into an afternoon task.

E-commerce catalogues grow faster than merchandising teams can tag them. Missing or inconsistent tags mean products do not appear in filtered searches, costing retailers 10-15% of potential revenue from undiscoverable inventory. GPU-accelerated tagging with open-source models brings consistency across your entire catalogue while adapting to new product lines without retraining.

Architecture Overview

The pipeline combines vision and language models. A vision model analyses product images to identify visual attributes — colour, pattern, shape, material appearance, and style. An LLM through vLLM processes product titles, descriptions, and extracted visual features to assign structured tags: category path, attribute key-value pairs, search keywords, and taxonomy codes.

The system uses your existing taxonomy as a reference, so generated tags conform to your category tree rather than inventing new structures. A validation step checks tag consistency — ensuring a product tagged as “leather” in material is not simultaneously tagged “vegan” in the lifestyle category, for example. Batch output integrates with your PIM or e-commerce platform via CSV or API.

GPU Requirements

Text-only tagging from descriptions runs significantly faster than combined vision-plus-text analysis. If your products have detailed descriptions, the text path alone achieves 90%+ accuracy. Adding image analysis pushes accuracy above 95% for visual attributes. See our self-hosted LLM guide for multi-modal model options.

Step-by-Step Build

Deploy vLLM on your GPU server with a model that handles structured extraction well. Prepare your taxonomy as a reference file. Build the tagging pipeline with batch processing.

# Product tagging prompt
TAG_PROMPT = """Tag this product using the provided taxonomy.
Product title: {title}
Description: {description}
Image analysis: {vision_output}

Taxonomy categories:
{taxonomy_tree}

Return JSON:
{category_path: ["Level1", "Level2", "Level3"],
 attributes: {colour: string, material: string, size_type: string,
              pattern: string, style: string},
 search_tags: [string],  # 8-12 relevant search terms
 taxonomy_code: string,
 confidence: 0.0-1.0}"""

# Batch processing
import pandas as pd, requests

catalogue = pd.read_csv("products.csv")
results = []
for _, row in catalogue.iterrows():
    prompt = TAG_PROMPT.format(
        title=row["title"], description=row["description"],
        vision_output=row.get("image_features", ""),
        taxonomy_tree=TAXONOMY
    )
    resp = requests.post(VLLM_URL, json={
        "model": MODEL, "messages": [{"role": "user", "content": prompt}],
        "max_tokens": 300, "temperature": 0.1
    })
    results.append(resp.json()["choices"][0]["message"]["content"])

For catalogues with images, run a vision model first to extract visual features, then pass those features alongside text to the tagging LLM. The OpenAI-compatible API makes it straightforward to swap models as better options emerge. See vLLM production setup for batch throughput tuning.

Quality and Consistency

Run the tagger on a sample of already-tagged products to measure accuracy against your ground truth. Typical accuracy exceeds 92% for category assignment and 88% for attribute extraction on first pass. Feed misclassified examples back into the prompt as few-shot examples to improve edge cases without model retraining.

Schedule nightly re-tagging runs for newly added products and periodic full-catalogue sweeps to catch taxonomy updates. Export audit reports showing tag distribution across categories to spot gaps in your product hierarchy.

Deploy Your Tagging Pipeline

Automated product tagging eliminates the merchandising bottleneck, ensures catalogue consistency, and improves product discoverability across search and filters. Process your entire catalogue on private infrastructure with no per-product API fees. Launch on GigaGPU dedicated GPU hosting and bring your catalogue under control. Explore more automation use cases and tutorials in our library.