LLM Models

DUTO uses large language models (LLMs) via OpenRouter for intelligence features like the Brain node, prompt expansion, and content analysis. The default model for all LLM operations is Gemini 2.5 Flash.

Available Models

Gemini 2.5 Flash

Fast, efficient reasoning

google/gemini-2.5-flash

Attribute	Value
Speed	Very Fast
Quality	Good
Context	Large
Cost	Low
Provider	OpenRouter

Strengths:

• Very fast responses
• Cost efficient
• Large context window
• Vision capable
• Good for simple tasks

Best for:

• Quick analysis
• Simple transformations
• High-volume processing
• Speed-critical tasks
• Image analysis with text

---

Gemini 2.5 Pro

High-quality reasoning

google/gemini-2.5-pro

Attribute	Value
Speed	Medium
Quality	Excellent
Context	Very Large
Cost	Medium
Provider	OpenRouter

Strengths:

• Excellent reasoning
• Complex task handling
• Nuanced understanding
• Creative capabilities
• Vision capable

Best for:

• Complex analysis
• Creative writing
• Detailed reasoning
• Important decisions
• Image-heavy analysis

---

DeepSeek

Technical and analytical model

deepseek/deepseek-r1

Attribute	Value
Speed	Medium
Quality	High
Context	Large
Cost	Low
Provider	OpenRouter

Strengths:

• Strong technical reasoning
• Good at structured tasks
• Cost effective
• Logical analysis
• Good for code-like tasks

Best for:

• Technical analysis
• Data processing
• Structured outputs
• Logic-heavy tasks

---

Model Comparison

Model	Speed	Quality	Cost	Best For
Gemini Flash	★★★★★	★★★	Low	Quick tasks
Gemini Pro	★★★	★★★★★	Medium	Quality tasks
DeepSeek	★★★★	★★★★	Low	Technical tasks

Usage in DUTO

Brain Node Modes

The Brain node uses LLMs differently based on mode:

Mode	Default Model	Use Case
Creative	Gemini 2.5 Flash	Creative writing, ideas
Research	Gemini 2.5 Flash	Analysis, synthesis (with tool calling)
Hybrid	Gemini 2.5 Flash	Balanced tasks (with tool calling)
Analysis	Gemini 2.5 Flash	Technical analysis

Research and Hybrid modes enable tool calling (web search, image search via Tavily), while Creative and Analysis modes use direct generation without tools.

Prompt Expander

The Prompt Expander node uses LLMs to enhance prompts:

Mode	Default Model	Purpose
Cinematic	Gemini Flash	Film-like descriptions
Commercial	Gemini Flash	Advertising style
Anime	Gemini Flash	Anime style
Documentary	Gemini Flash	Realistic style

Content Analysis

LLMs are used for:

• Text classification
• Sentiment analysis
• Content summarization
• Image description
• Reference analysis

OpenRouter Integration

DUTO uses OpenRouter as its LLM gateway, providing access to multiple model providers through a single API:

Edge Functions:

• brain-reasoning - Brain node agentic reasoning (uses Vercel AI SDK)
• openrouter-chat - Prompt expansion, analyzer, and general LLM tasks

Default Model: google/gemini-2.5-flash (used for both fast and creative tasks)

Available Models:

• google/gemini-2.5-flash - Default, fast, cost-efficient
• google/gemini-2.5-pro - Higher quality reasoning

Credit Costs

Model	Simple Query	Complex Task
Gemini Flash	1 credit	2 credits
Gemini Pro	2-3 credits	4-5 credits
DeepSeek	1-2 credits	2-3 credits

Task Recommendations

Simple Tasks

Use Gemini Flash for:

• Extracting data
• Simple transformations
• Quick categorization
• Format conversion
• Basic prompt expansion

Complex Tasks

Use Gemini Pro for:

• Creative writing
• Complex analysis
• Nuanced decisions
• Multi-step reasoning
• Image-heavy analysis
• Story generation

Technical Tasks

Use DeepSeek for:

• Data parsing
• Structured extraction
• Technical analysis
• Logic-heavy tasks
• Code generation

Prompt Optimization

For All Models

1. Be specific

   ✓ "Extract the product name and price from this description"
   ✗ "Get info from this"

2. Provide context

   ✓ "This is a product description. Extract: name, price, category"
   ✗ "Parse this text"

3. Specify format

   ✓ "Return as JSON: {name: string, price: number}"
   ✗ "Give me the data"

Gemini Flash Optimization

Keep prompts concise:

Extract product name and price as JSON.
Input: [text]
Output: {"name": "...", "price": ...}

Gemini Pro Optimization

Can handle complex prompts:

Analyze this product description and provide:
1. Product name
2. Key features (list)
3. Target audience
4. Suggested improvements

Consider market positioning and competitor analysis.

DeepSeek Optimization

Structure requests clearly:

Task: Parse the following data
Format: JSON array
Fields: name, email, company
Rules:
- Normalize email to lowercase
- Extract company from domain if not stated

Advanced Usage

Chain of Thought

For complex reasoning:

Think through this step by step:
1. First, identify...
2. Then, analyze...
3. Finally, conclude...

Output Schemas

Define exact output structure:

{
  "analysis": {
    "sentiment": "positive|negative|neutral",
    "confidence": 0.0-1.0,
    "key_points": ["string"]
  }
}

Few-Shot Examples

Provide examples for consistency:

Example input: "Blue cotton t-shirt, size M, $29.99"
Example output: {"product": "t-shirt", "color": "blue", "price": 29.99}

Now process: "Red wool sweater, size L, $49.99"

Troubleshooting

Inconsistent Outputs

Problem: Same input gives different outputs

Solutions:

1. Be more specific in prompt
2. Use output schema
3. Provide examples
4. Use Pro model for consistency

Wrong Format

Problem: Output doesn't match expected format

Solutions:

1. Specify format explicitly
2. Provide JSON schema
3. Include example output
4. Parse and validate in workflow

Slow Responses

Problem: LLM taking too long

Solutions:

1. Use Flash model
2. Simplify prompt
3. Reduce output requirements
4. Break into smaller tasks

Vision Capabilities

Both Gemini models support vision input for:

• Image analysis
• Reference description
• Visual Q&A
• Content moderation
• Storyboard context

When using image inputs with LLMs:

• Provide clear context for the image
• Specify what to extract/analyze
• Use appropriate model (Pro for complex analysis)
• Consider image size/quality