LLM Models

Sellestial supports multiple Large Language Models (LLMs) across different pipeline kinds.

Available Models by Pipeline Kind

Classifier

Available models for classification pipelines:

Model	Provider	Best For
`openai/o1`	OpenAI	Complex reasoning
`openai/o4-mini`	OpenAI	Fast, economical
`openai/o3`	OpenAI	Balanced performance
`openai/gpt-4.1`	OpenAI	High accuracy
`openai/gpt-4.1-mini`	OpenAI	Balanced
`openai/gpt-4.1-nano`	OpenAI	Ultra-fast
`gemini/gemini-2.5-pro`	Google	Premium quality
`gemini/gemini-2.0-flash`	Google	Speed and cost

StructuredData

Available models for structured data extraction:

Model	Provider	Best For
`openai/gpt-5`	OpenAI	Latest capabilities
`openai/gpt-oss-120b`	OpenAI	Open-source model
`openai/o1`	OpenAI	Complex reasoning
`openai/o4-mini`	OpenAI	Fast, economical
`openai/o3`	OpenAI	Balanced performance
`openai/gpt-4.1`	OpenAI	High accuracy
`openai/gpt-4.1-mini`	OpenAI	Balanced
`openai/gpt-4.1-nano`	OpenAI	Ultra-fast
`gemini/gemini-2.5-pro`	Google	Premium quality
`gemini/gemini-2.0-flash`	Google	Speed and cost
`gemini/gemini-2.5-flash`	Google	Fast and efficient

Agent

Available models for agent pipelines (generation and validation):

Model	Provider	Best For
`openai/o1`	OpenAI	Complex reasoning
`openai/o4-mini`	OpenAI	Fast, economical
`openai/o3`	OpenAI	Balanced performance (recommended for validation)
`gemini/gemini-2.0-flash`	Google	Speed and cost (recommended for generation)

Sequence

Available models for sequence content generation:

Model	Provider	Best For
`openai/o1`	OpenAI	Complex reasoning
`deepseek/r1`	DeepSeek	Technical content
`openai/o4-mini`	OpenAI	Fast, economical
`openai/o3`	OpenAI	Balanced performance
`openai/gpt-4.1`	OpenAI	Quality generation
`gemini/gemini-2.5-pro-preview-05-06`	Google	Preview features
`gemini/gemini-2.5-pro`	Google	Premium quality
`gemini/gemini-2.0-flash`	Google	Speed and cost

Agent Configuration

Agent Model Configuration

Generation Model:

Primary model for reasoning and decisions
Default: gemini/gemini-2.0-flash

Validation Model:

Reviews and validates outputs
Default: openai/o3 or openai/o4-mini
Optional: Can disable validation

Model Characteristics

OpenAI o-series

openai/o1:

Complex multi-step reasoning
Highest quality outputs
Slower processing
Higher cost
Best for: Critical decisions, complex analysis

openai/o3:

Balanced quality and speed
Good reasoning capabilities
Moderate cost
Best for: General purpose, most use cases

openai/o4-mini:

Fast processing
Lower cost
Good for simple tasks
Best for: High-volume operations, simple classification

OpenAI GPT-4.1 series

openai/gpt-4.1:

High-quality generation
Strong reasoning
Content creation
Best for: Sequences, content generation

openai/gpt-4.1-mini:

Faster than gpt-4.1
Lower cost
Good quality
Best for: Classifiers, simple generation

openai/gpt-4.1-nano:

Ultra-fast
Very low cost
Basic capabilities
Best for: Simple classification, high-volume

Google Gemini series

gemini/gemini-2.0-flash:

Very fast
Cost-effective
Good quality
Best for: Default choice for most pipelines

gemini/gemini-2.5-pro:

Premium quality
Advanced reasoning
Multimodal capabilities
Best for: Complex tasks, high-value operations

gemini/gemini-2.5-pro-preview:

Latest features
Experimental capabilities
May change
Best for: Testing new features

DeepSeek

deepseek/r1:

Technical content
Code generation
Mathematical reasoning
Best for: Technical sequences, developer content

Choosing a Model

By Use Case

Data Classification:

Start with: openai/o4-mini
Upgrade to: openai/gpt-4.1-mini for better accuracy
Premium: gemini/gemini-2.5-pro

Structured Data Extraction:

Start with: openai/o4-mini or gemini/gemini-2.0-flash
Upgrade to: openai/gpt-4.1 for complex extraction
Latest: openai/gpt-5 for cutting-edge capabilities

Complex Analysis (Agents):

Generation: gemini/gemini-2.0-flash
Validation: openai/o3

Content Generation (Sequences):

High quality: gemini/gemini-2.5-pro
Balanced: openai/gpt-4.1
Fast: gemini/gemini-2.0-flash

Simple Logic (Code):

No AI model needed (pure code)

By Priority

Quality First:

gemini/gemini-2.5-pro
openai/o1
openai/gpt-4.1

Speed First:

openai/o4-mini
gemini/gemini-2.0-flash
openai/gpt-4.1-nano

Cost First:

openai/o4-mini
openai/gpt-4.1-nano
gemini/gemini-2.0-flash

Balanced:

openai/o3
gemini/gemini-2.0-flash
openai/gpt-4.1-mini

By Volume

High Volume (greater than 1000 records/day):

openai/o4-mini — Fast and economical
gemini/gemini-2.0-flash — Good balance

Medium Volume (100-1000 records/day):

openai/o3 — Balanced quality
gemini/gemini-2.5-pro — If quality critical

Low Volume (less than 100 records/day):

Any model works
Choose based on quality needs

Model Configuration

In Pipelines

Classifier:

kind: Classifier
spec:
  classifier:
    llmModel: openai/o4-mini

StructuredData:

kind: StructuredData
spec:
  structuredData:
    llmModel: openai/gpt-4.1

Agent:

kind: Agent
spec:
  agent:
    agentDef:
      llmModel: gemini/gemini-2.0-flash
      validationLlmModel: openai/o3
      enableValidation: true

Sequence:

kind: Sequence
spec:
  llmModel: gemini/gemini-2.5-pro

Changing Models

Go to pipeline Generator/Classifier/Agent/StructuredData tab
Find LLM Model selector
Choose from dropdown
Save changes
Compile and deploy new version

Cost Considerations

Relative Costs

(Approximate, normalized)

Model	Relative Cost
openai/o4-mini	1x (baseline)
openai/gpt-4.1-nano	1x
gemini/gemini-2.0-flash	1.5x
openai/o3	3x
openai/gpt-4.1-mini	4x
openai/gpt-4.1	8x
gemini/gemini-2.5-pro	10x
openai/o1	15x

Cost Optimization

For Classifiers:

Start with: openai/o4-mini (1x cost)
Test accuracy
If insufficient: openai/gpt-4.1-mini (4x cost)

For Agents:

Generation: gemini/gemini-2.0-flash (1.5x)
Validation: openai/o3 (3x)
Or disable validation for lower cost

For Sequences:

Testing: gemini/gemini-2.0-flash (1.5x)
Production: gemini/gemini-2.5-pro (10x) if quality critical
Or: openai/gpt-4.1 (8x) for balance

Performance Characteristics

Speed

Fastest:

openai/o4-mini
openai/gpt-4.1-nano
gemini/gemini-2.0-flash

Moderate:

openai/o3
openai/gpt-4.1-mini
deepseek/r1

Slower:

openai/gpt-4.1
gemini/gemini-2.5-pro
openai/o1

Quality

Highest:

openai/o1
gemini/gemini-2.5-pro
openai/gpt-4.1

Good:

openai/o3
openai/gpt-4.1-mini
gemini/gemini-2.0-flash

Basic:

openai/o4-mini
openai/gpt-4.1-nano

Validation Models (Agents Only)

Agents can use a separate validation model to review outputs.

Purpose

Catch errors in agent output
Enforce schema compliance
Verify logic consistency
Provide feedback for corrections

Common Patterns

High-Quality Validation:

Generation: gemini/gemini-2.0-flash
Validation: openai/o3

Cost-Optimized:

Generation: gemini/gemini-2.0-flash
Validation: openai/o4-mini

Premium:

Generation: gemini/gemini-2.5-pro
Validation: openai/o1

When to Disable

Disable validation if:

Simple, well-defined outputs
Cost is a major concern
Speed is critical
Outputs rarely have errors

Keep enabled for:

Complex decision-making
Critical operations (merges, deletions)
Structured output requirements
High-value records

Best Practices

Model Selection

Testing

Test methodology:

Select model
Process 10-20 sample records
Review outputs manually
Check quality, accuracy, tone
If insufficient, try next tier model
Compare cost vs. quality benefit

Monitoring

Track:

Success rates
Error rates
Cost per record
Processing time

Adjust if:

High error rates → Upgrade model
Slow processing → Use faster model
High costs → Downgrade if quality acceptable

Troubleshooting

Model Not Available

Problem: Model not in dropdown

Solutions:

Check pipeline kind compatibility
Verify account has access
Try alternate model
Contact support

Poor Quality Outputs

Problem: Model producing bad results

Solutions:

Review prompt quality
Check input data completeness
Try higher-tier model
Add validation (agents)
Improve prompt engineering

High Costs

Problem: Model too expensive

Solutions:

Use faster, cheaper model
Process smaller batches
Filter inputs before processing
Disable validation if not needed
Run only on high-value records

Next Steps

Pipeline Kinds — Choose the right pipeline kind
Data Sources — Available data providers
Best Practices — Optimization tips