Build a Bulk Product Description Generator That Actually Ships (Python + GPT-4) published: false

The Mango Wood Serving Board brings warmth and texture to any table setting.
Hand-finished from sustainably sourced mango wood, each board carries a 
naturally unique grain pattern — no two are identical. At 14x10 inches, it's 
sized for charcuterie spreads, cheese boards, or everyday serving. The juice 
groove keeps counters clean. The integrated handle makes it easy to carry 
from kitchen to table without a second trip.

Dimensions: 14" x 10" x 0.75" | Material: Mango wood | Care: Hand wash only

pip install openai pandas tenacity

product_name, material, dimensions, features, category

CSV Input
   ↓
Load & validate rows
   ↓
Check progress file (skip already-processed items)
   ↓
For each product:
  → Build prompt
  → Call GPT-4o (with retry logic)
  → Validate output
  → Write to progress file
   ↓
Merge all outputs → final CSV

def build_prompt(row: dict) -> str:
    return f"""You are a product copywriter for a premium e-commerce brand.

Write a product description for the item below. Follow these rules exactly:

1. Start with a single sentence that leads with the product's primary use or benefit — not its name.
2. Write 3-4 sentences of body copy. Focus on what the customer will *experience*, not just what the product *is*.
3. End with a one-line spec summary: Dimensions, Material, and one care/compatibility note.
4. Tone: warm, specific, confidence-inspiring. No filler phrases like "perfect for" or "great for any occasion."
5. Total length: 80-120 words. Not shorter. Not longer.

Product data:
- Name: {row['product_name']}
- Material: {row['material']}
- Dimensions: {row['dimensions']}
- Features: {row['features']}
- Category: {row['category']}

Output the description only. No intro, no labels, no quotes around the text."""

import openai
from tenacity import (
    retry,
    stop_after_attempt,
    wait_exponential,
    retry_if_exception_type
)

client = openai.OpenAI(api_key="your-api-key-here")

@retry(
    retry=retry_if_exception_type(openai.RateLimitError),
    wait=wait_exponential(multiplier=1, min=2, max=60),
    stop=stop_after_attempt(6)
)
def generate_description(prompt: str) -> str:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}],
        max_tokens=300,
        temperature=0.7,  # higher = more varied output; lower = more consistent
    )
    return response.choices[0].message.content.strip()

def validate_description(text: str, product_name: str) -> tuple[bool, str]:
    word_count = len(text.split())

    # Too short — model probably got confused or hit an error mid-generation
    if word_count < 60:
        return False, f"Too short: {word_count} words"

    # Too long — prompt constraint wasn't followed
    if word_count > 150:
        return False, f"Too long: {word_count} words"

    # Check that the product name appears (basic relevance check)
    # Use the first significant word of the product name only
    key_term = product_name.split()[0].lower()
    if key_term not in text.lower() and len(product_name.split()) > 1:
        # Partial match is fine — GPT-4o often refers to "the board" not "the Mango Wood Serving Board"
        pass

    # Catch GPT-4o's habit of adding its own preamble
    bad_prefixes = ["here's", "here is", "certainly", "of course", "sure,"]
    first_words = text[:30].lower()
    for prefix in bad_prefixes:
        if first_words.startswith(prefix):
            return False, f"Model preamble detected: starts with '{prefix}'"

    return True, "ok"

import json
import os

PROGRESS_FILE = "progress.json"

def load_progress() -> dict:
    if os.path.exists(PROGRESS_FILE):
        with open(PROGRESS_FILE, "r") as f:
            return json.load(f)
    return {}

def save_progress(progress: dict):
    with open(PROGRESS_FILE, "w") as f:
        json.dump(progress, f, indent=2)

progress = load_progress()

for _, row in df.iterrows():
    product_id = str(row['product_name'])  # or use a SKU if you have one

    # Skip if already processed
    if product_id in progress:
        print(f"Skipping {product_id} — already done")
        continue

    prompt = build_prompt(row.to_dict())

    try:
        description = generate_description(prompt)
        is_valid, reason = validate_description(description, row['product_name'])

        if not is_valid:
            print(f"FAILED validation for {product_id}: {reason}")
            # Still save as failed so we can review separately
            progress[product_id] = {"status": "failed", "reason": reason, "text": description}
        else:
            progress[product_id] = {"status": "ok", "text": description}

        save_progress(progress)  # write after every item

    except Exception as e:
        print(f"ERROR on {product_id}: {e}")
        # Don't save — we'll retry this one next run

import pandas as pd
import json

def run_pipeline(input_csv: str, output_csv: str):
    df = pd.read_csv(input_csv)
    progress = load_progress()

    total = len(df)
    processed = 0
    failed = 0

    print(f"Starting pipeline: {total} products, {len(progress)} already processed")

    for _, row in df.iterrows():
        product_id = str(row['product_name'])

        if product_id in progress and progress[product_id]['status'] == 'ok':
            processed += 1
            continue

        prompt = build_prompt(row.to_dict())

        try:
            description = generate_description(prompt)
            is_valid, reason = validate_description(description, row['product_name'])

            status = "ok" if is_valid else "failed"
            progress[product_id] = {
                "status": status,
                "text": description,
                "reason": reason if not is_valid else None
            }
            save_progress(progress)

            if is_valid:
                processed += 1
                print(f"✓ {product_id} ({processed}/{total})")
            else:
                failed += 1
                print(f"✗ {product_id} — {reason}")

        except Exception as e:
            failed += 1
            print(f"ERROR: {product_id} — {e}")

    # Build output CSV from progress file
    results = []
    for _, row in df.iterrows():
        product_id = str(row['product_name'])
        entry = progress.get(product_id, {})
        results.append({
            **row.to_dict(),
            "description": entry.get("text", ""),
            "status": entry.get("status", "missing")
        })

    output_df = pd.DataFrame(results)
    output_df.to_csv(output_csv, index=False)
    print(f"\nDone. {processed} succeeded, {failed} failed.")
    print(f"Output: {output_csv}")

# Run it
run_pipeline("products.csv", "products_with_descriptions.csv")

df = pd.read_csv(input_csv, encoding='utf-8-sig')
# or
df = pd.read_csv(input_csv, encoding='latin-1')

"Avoid phrases like: 'perfect for', 'great for any occasion', 'ideal for the whole family', 'make a statement', 'elevate your space'."