E-Commerce Product Catalog AI Classification: Fine-Tuned Category Models

E-commerce brands adding 100-500 new SKUs per month face a catalog management problem: every new product needs to be categorized, tagged, attributed, and placed in the right navigation structure. Done manually, this takes 5-15 minutes per product — 8-75 hours per month in direct labor.

A fine-tuned classifier trained on your taxonomy does it in seconds per product, at 90%+ accuracy. This is a straightforward AI agency deliverable: clear before/after metrics, fast build time, and an obvious retainer justification (new products come in every month).

What the Classifier Does

Input: Product data (name, description, brand, any existing attributes)

Output: Classification across multiple dimensions:

Primary category (Clothing > Men's > Outerwear)
Secondary tags (waterproof, insulated, packable)
Gender/size range
Material classification
Price tier
Search keywords

The model outputs structured JSON that your catalog management system ingests directly.

Example:

Input:

Product: Arc'teryx Beta AR Jacket Men's
Description: All-round waterproof shell for mountain activities. GORE-TEX Pro fabric, fully seam-taped, helmet-compatible hood. 485g.

Output:

{
  "primary_category": "Clothing > Men's > Jackets & Coats > Rain Jackets",
  "secondary_categories": ["Hiking", "Mountaineering", "Skiing"],
  "attributes": {
    "waterproof": true,
    "material": "GORE-TEX Pro",
    "insulation": "none",
    "gender": "mens",
    "weight_oz": 17.1,
    "packable": true
  },
  "tags": ["waterproof", "shell", "gore-tex", "mountaineering", "packable", "alpine"],
  "price_tier": "premium",
  "meta_keywords": ["waterproof jacket mens", "gore-tex jacket", "mountain shell", "rain jacket hiking"]
}

Why a Fine-Tuned Model Outperforms Generic AI

Generic GPT-4 with a prompt can classify products at a basic level. The problems:

It does not know your taxonomy. Your store has a specific category structure with 3-4 levels. Generic AI invents categories that do not exist in your navigation.
It does not know your attribute vocabulary. Your "price tier" definitions, your material classifications, your activity tags — these are store-specific. Generic AI guesses.
It is not calibrated to your edge cases. A down jacket with a waterproof shell is in what category? A swimsuit for men sold in an outdoor sports store — swimwear or outdoor gear? Your past catalog decisions encode your answer; a fine-tuned model learns it.

Building the Dataset

Source: Your existing classified product catalog — every product you have already categorized manually is a training example.

Size target: 1,000-5,000 products (covering your category range)

Construction:

{"messages": [
  {"role": "system", "content": "You are a product classification assistant for [Brand]. Classify products according to our taxonomy. Always output valid JSON matching the schema provided."},
  {"role": "user", "content": "Classify this product:\nName: Patagonia Nano Puff Jacket Womens\nDescription: Lightweight insulated jacket with PrimaLoft Gold Insulation Eco. Wind resistant DWR finish. Packs into chest pocket. 9.5 oz."},
  {"role": "assistant", "content": "{\"primary_category\": \"Clothing > Women's > Jackets & Coats > Insulated Jackets\", \"secondary_categories\": [\"Hiking\", \"Travel\", \"Skiing\"], \"attributes\": {\"waterproof\": false, \"material\": \"PrimaLoft Gold\", \"insulation\": \"synthetic\", \"gender\": \"womens\", \"weight_oz\": 9.5, \"packable\": true}, \"tags\": [\"insulated\", \"packable\", \"lightweight\", \"synthetic-fill\", \"primaloft\"], \"price_tier\": \"premium\", \"meta_keywords\": [\"insulated jacket women\", \"packable down jacket\", \"lightweight insulated jacket\"]}"}
]}

Include examples from every category in your taxonomy. Aim for 20-50 examples per top-level category.

Training Configuration

For classification tasks with structured JSON output:

Base model: Mistral 7B Instruct performs well on structured output tasks
LoRA rank: 8-16 (lower rank is fine for classification)
Epochs: 3-5 (classification tasks converge quickly)

The model needs to learn: (1) your category structure, (2) your attribute vocabulary, (3) how to output valid JSON.

Evaluation

Hold out 10% of your dataset. After training, run the evaluation set and measure:

Primary metric: Correct primary category assignment (exact match)

Secondary metrics:

Tag precision (tags assigned that are correct)
Tag recall (correct tags that were assigned)
JSON validity (100% of outputs should be parseable)
Attribute accuracy (individual field accuracy)

Typical results with a well-constructed 2,000+ example dataset: 88-94% correct primary category on held-out set.

Integration

Batch classification pipeline for new product ingestion:

import requests
import json

def classify_product(name: str, description: str) -> dict:
    response = requests.post(
        'http://your-ollama-server:11434/api/chat',
        json={
            "model": "product-classifier",
            "messages": [
                {
                    "role": "user",
                    "content": f"Classify this product:\nName: {name}\nDescription: {description}"
                }
            ],
            "stream": False
        }
    )

    content = response.json()['message']['content']

    try:
        return json.loads(content)
    except json.JSONDecodeError:
        # Extract JSON from response if wrapped in text
        import re
        json_match = re.search(r'\{.*\}', content, re.DOTALL)
        if json_match:
            return json.loads(json_match.group())
        raise ValueError(f"Could not parse classification output: {content}")

# Process new products CSV
import csv
with open('new_products.csv') as f:
    for row in csv.DictReader(f):
        classification = classify_product(row['name'], row['description'])
        # Push to your catalog management system
        update_catalog(row['sku'], classification)

Run this as a nightly job on new product imports. Agent review catches the 6-12% that need manual correction.

Retainer Structure for This Use Case

The retainer for catalog classification is justified by:

New products arrive continuously → model processes them automatically
Taxonomy changes (new categories, restructured navigation) → model needs retraining
Accuracy monitoring → catching classification drift before it pollutes your catalog

Retainer package: $300-500/month

Includes: Monthly batch processing of new products, quarterly retraining with new examples, accuracy monitoring dashboard, corrections pipeline for agent feedback