AI-Powered Content Generation: DeepSeek for Product Descriptions

This article explains how we use DeepSeek AI to generate product descriptions, translate content, and create marketing copy at scale.

The Problem: Scaling Content Creation

We have 5,000+ product configurations that need:

Product descriptions: 200-300 words per product
Meta descriptions: 150-160 characters for SEO
Feature highlights: Bullet points for key features
Marketing copy: Compelling sales language

Writing manually:

5,000 products × 30 minutes per product = 2,500 hours (104 days)
Consistency: Ensure all descriptions follow brand voice
Updates: Rewrite when products change
Translations: Multiply by 10 languages = 25,000 hours

This doesn't scale. We need AI automation.

The Solution: DeepSeek AI

We use DeepSeek AI model to generate content:

Model: DeepSeek-V3 (671B parameters)
API: deepseek.com with Together.ai fallback
Cost: $0.27 per million input tokens, $1.10 per million output tokens
Speed: ~50 tokens/second (10 seconds per description)

This enables generating 5,000 descriptions in ~14 hours at ~$50 cost.

Unified AI Generation Interface

We built a unified interface supporting multiple providers:

from app.shared.ai_generation import generate_content

description = generate_content(
    prompt="Write a product description for Treo N100 Mini PC",
    provider="deepseek",
    max_tokens=800,
    temperature=0.7,
    system_prompt="You are a technical writer..."
)

Supported Providers:

DeepSeek: Primary provider (deepseek.com)
Gemini: Fallback provider (Google)
Together.ai: DeepSeek fallback (if deepseek.com fails)

DeepSeek API Integration

Direct API Call

We call deepseek.com API directly:

response = requests.post(
    "https://api.deepseek.com/v1/chat/completions",
    headers={
# ... (implementation details omitted)

Fallback to Together.ai

If deepseek.com fails, we fall back to Together.ai:

response = requests.post(
    "https://api.together.xyz/v1/chat/completions",
    headers={
# ... (implementation details omitted)

This ensures high availability even if one provider is down.

System Prompt Design for Caching

DeepSeek supports prompt caching to reduce costs:

System prompt: Cached (reused across requests)
User prompt: Not cached (unique per request)

We design system prompts to be reusable:

system_prompt = """You are a technical writer for Thinvent, a manufacturer of mini PCs, thin clients, and industrial computers.

Write product descriptions that:
# ... (implementation details omitted)

This system prompt is cached and reused for all 5,000 products, saving ~$40 per run.

Cache Hit Logging

We log cache hits to monitor savings:

usage = result.get("usage", {})
cache_hits = usage.get("prompt_cache_hit_tokens", 0)
cache_misses = usage.get("prompt_cache_miss_tokens", 0)

if cache_hits > 0:
    logging.info(f"DeepSeek cache HIT: {cache_hits} tokens (miss: {cache_misses})")

Example output:

DeepSeek cache HIT: 1,200 tokens (miss: 50)
DeepSeek cache HIT: 1,200 tokens (miss: 48)
DeepSeek cache HIT: 1,200 tokens (miss: 52)

This shows the system prompt (1,200 tokens) is cached, while user prompts (50 tokens) are not.

Temperature Settings

We use different temperatures for different use cases:

Product Descriptions (0.7)

Purpose: Balanced creativity and consistency

description = generate_content(
    prompt=f"Write a description for {product_name}",
    temperature=0.7
)

Result: Natural-sounding descriptions with some variation

Translations (1.3)

Purpose: More creative, natural translations

translation = generate_content(
    prompt=f"Translate to {lang}: {text}",
    temperature=1.3
)

Result: Idiomatic translations that sound native

Technical Specs (0.3)

Purpose: Factual, consistent output

specs = generate_content(
    prompt=f"List specifications for {product_name}",
    temperature=0.3
)

Result: Consistent, accurate specifications

Marketing Copy (1.0)

Purpose: Creative, engaging content

copy = generate_content(
    prompt=f"Write marketing copy for {product_name}",
    temperature=1.0
)

Result: Compelling, varied marketing language

Use Cases

1. Product Descriptions

Input: Product SKU and features
Prompt:

Write a product description for:

- Name: Treo N100 Mini PC

- Processor: Intel N100 (4 cores, 3.4GHz)

- RAM: 8GB DDR4

- Storage: 256GB SSD

- Ports: 2× HDMI, 4× USB 3.0, Gigabit Ethernet

- OS: Windows 11 Pro

Output:

The Thinvent Treo N100 Mini PC is a compact powerhouse designed for business and industrial applications. Powered by the Intel N100 processor with 4 cores running at 3.4GHz, it delivers reliable performance for everyday computing tasks.

With 8GB DDR4 RAM and a 256GB SSD, the Treo provides smooth multitasking and fast boot times. The dual HDMI ports support multi-monitor setups, while 4 USB 3.0 ports and Gigabit Ethernet ensure seamless connectivity.

Pre-installed with Windows 11 Pro, the Treo is ready for business use out of the box. Its fanless design ensures silent operation, making it ideal for office environments, digital signage, and industrial automation.

Backed by Thinvent's 3-year warranty and certified to BIS, RoHS, and ISO standards, the Treo N100 Mini PC is built to last.

2. Content Translation

Input: English text and target language
Prompt:

Translate to Hindi:
"The Thinvent Treo N100 Mini PC is a compact powerhouse designed for business and industrial applications."

Output:

Thinvent Treo N100 मिनी पीसी एक कॉम्पैक्ट पावरहाउस है जो व्यावसायिक और औद्योगिक अनुप्रयोगों के लिए डिज़ाइन किया गया है।

3. Meta Descriptions

Input: Product name and key features
Prompt:

Write a 150-character meta description for:
Treo N100 Mini PC with Intel N100, 8GB RAM, 256GB SSD, Windows 11 Pro

Output:

Compact Treo N100 Mini PC with Intel N100, 8GB RAM, 256GB SSD, Windows 11 Pro. Perfect for business and industrial use. 3-year warranty.

4. Feature Highlights

Input: Product features
Prompt:

Write 5 bullet points highlighting key features of:
Treo N100 Mini PC

Output:

• Intel N100 processor (4 cores, 3.4GHz) for reliable performance
• 8GB DDR4 RAM and 256GB SSD for smooth multitasking
• Dual HDMI ports for multi-monitor setups
• Fanless design for silent operation
• Windows 11 Pro pre-installed, ready for business use

Batch Processing

We process content in batches for efficiency:

from app.web.models_product_description import ProductDescription

# Queue descriptions for generation
# ... (implementation details omitted)

This allows processing thousands of descriptions in the background.

Error Handling

We handle API failures gracefully:

try:
    content = generate_content(prompt, provider="deepseek")
    if content:
# ... (implementation details omitted)

If generation fails, we: 1. Log the error 2. Return None (caller handles fallback) 3. Retry later (for queued items)

Cost Optimization

We optimize costs through:

1. Prompt Caching

System prompt: 1,200 tokens × 5,000 products = 6M tokens
Without caching: $0.27 × 6 = $1.62
With caching: $0.27 × 0.006 = $0.0016 (99.9% savings)

2. Batch Processing

Process multiple items in one API call:

prompt = "Write descriptions for:\n"
for sku in batch_skus:
    prompt += f"- {sku}\n"

This reduces API overhead and increases throughput.

3. Temperature Tuning

Lower temperature = fewer tokens generated:

Temperature 0.3: ~200 tokens per description
Temperature 0.7: ~250 tokens per description
Temperature 1.3: ~300 tokens per description

Use lower temperature when possible.

Performance Characteristics

Generation Speed:

Single description: ~10 seconds
Batch of 10: ~30 seconds (3 seconds per description)
5,000 descriptions: ~14 hours (batch processing)

Cost:

Input tokens: 1,200 (system) + 50 (user) = 1,250 tokens
Output tokens: 250 tokens
Cost per description: $0.27 × 0.00125 + $1.10 × 0.00025 = $0.00061
5,000 descriptions: $3.05

With caching:

Input tokens: 6 (cached) + 50 (user) = 56 tokens
Cost per description: $0.27 × 0.000056 + $1.10 × 0.00025 = $0.00029
5,000 descriptions: $1.45 (52% savings)

Integration with SEO Pipeline

AI-generated content integrates with the SEO pipeline:

Step 9: Migrate Descriptions

The SEO pipeline generates descriptions for all products:

for sku in product_skus:
    description = generate_content(
        prompt=f"Write description for {sku}",
# ... (implementation details omitted)

See SEO Pipeline Overview for details.

Translation Queue

Missing translations are queued and processed with AI:

for item in translation_queue:
    translation = generate_content(
        prompt=f"Translate to {item['lang']}: {item['text']}",
        provider="deepseek",
        temperature=1.3
    )

    translation_manager.set(item['text'], item['lang'], translation)

See Translation System for details.

References

AI Services

DeepSeek - Official website
DeepSeek API Documentation - API docs
Together.ai - Fallback provider

Technical Concepts

Prompt Caching - DeepSeek docs
Temperature Sampling - Wikipedia

Translation System - How we translate content
SEO Pipeline Overview - Complete pipeline architecture
System Prompt Design - Designing reusable prompts

Summary

We use DeepSeek AI to generate content at scale:

Model: DeepSeek-V3 (671B parameters)
API: deepseek.com with Together.ai fallback
Use Cases:
Product descriptions (0.7 temperature)
Translations (1.3 temperature)
Technical specs (0.3 temperature)
Marketing copy (1.0 temperature)

Optimization:

Prompt caching (99.9% savings on system prompt)
Batch processing (3× faster)
Temperature tuning (20% fewer tokens)

Performance:

5,000 descriptions in ~14 hours
Cost: $1.45 with caching (52% savings)
Speed: ~3 seconds per description (batched)

Integration:

SEO pipeline (Step 9)
Translation queue (background worker)
Product description table (DynamoDB)

This AI-powered approach enables us to generate and maintain thousands of product descriptions across multiple languages without manual effort.

← Back to Documentation Index

Products

Popular Searches and Blogs

AI-Powered Content Generation: DeepSeek for Product Descriptions

The Problem: Scaling Content Creation

The Solution: DeepSeek AI

Unified AI Generation Interface

DeepSeek API Integration

Direct API Call

Fallback to Together.ai

System Prompt Design for Caching

Cache Hit Logging

Temperature Settings

Product Descriptions (0.7)

Translations (1.3)

Technical Specs (0.3)

Marketing Copy (1.0)

Use Cases

1. Product Descriptions

2. Content Translation

3. Meta Descriptions

4. Feature Highlights

Batch Processing

Error Handling

Cost Optimization

1. Prompt Caching

2. Batch Processing

3. Temperature Tuning

Performance Characteristics

Integration with SEO Pipeline

Step 9: Migrate Descriptions

Translation Queue

References

AI Services

Technical Concepts

Related Articles

Summary