šÆ Overview#
The AI Block allows you to build a Prompt using one or several Embedded Blocks. This Prompt will generate dynamic content during the Publication completion by the User.By default, the AI Block allows you to add as many Embedded Blocks as necessary, similar to a Group Block, but the content of the Embedded Blocks will constitute the entire Prompt.It is therefore possible to break it down into several levels, make it complex, with variables and entire portions hidden or visible according to different conditional rules.
āļø Specifications#
It is essential to keep in mind that all the visible content in the AI Block, once the Attributes are assigned via the Assistant/Wizard, constitutes a Prompt. It is necessary to adapt the formulations to make the Prompt as effective as possible.If you need a very long prompt, adapt the model choice accordingly so it can handle it.šļø AI Model Selection#
From the AI Block settings, you can also choose the model to use.š Model Overview#
Thinkeo offers an extensive selection of AI models, each with its own reasoning capabilities and context handling capacity. It is important to adapt your model choice to the specific use of your application.Models present their context capacity in tokens. In French, we estimate that one word averages 1.5 tokens. The total context is shared between the input (prompt) and output (response) of a model. Each company offering AI models generally provides two alternatives: an intelligent model for complex tasks and another faster and less expensive one adapted to simple tasks.š§ o Series (Advanced Reasoning Models)
o4 MiniModel with reinforced reasoning capabilities
Efficiency-optimized version
Adapted for very complex reasoning
Total context: 128,000 tokens
Response length: up to 65,536 tokens
Advanced model with deep reasoning
Excellent performance on very complex reasoning tasks
Total context: 128,000 tokens
Response length: up to 100,000 tokens
Lightweight version of o3 with integrated reasoning
Economic alternative for complex reasoning
Total context: 128,000 tokens
Response length: up to 65,536 tokens
ā” GPT-5 Series
GPT-5OpenAI's most advanced model with unified reasoning and task execution
Automatically switches between fast responses and deep thinking when needed
Excellent for complex coding, writing, health, and analysis tasks
Total context: 400,000 tokens
Response length: up to 128,000 tokens
Lightweight version optimized for speed and cost efficiency
Maintains high performance for most common tasks
Ideal for applications requiring frequent API calls
Total context: 400,000 tokens
Response length: up to 128,000 tokens
Ultra-compact version for high-throughput, low-latency applications
Best performance-to-cost ratio for simple to medium complexity tasks
Perfect for real-time applications
Total context: 400,000 tokens
Response length: up to 128,000 tokens
šÆ GPT-4 Series
GPT-4.1 (Recommended by Thinkeo)Best performance/cost/versatility ratio
Excellent reasoning capabilities with massive context
Efficiently processes large documents while remaining economical
Total context: 1,000,000 tokens
Response length: up to 16,384 tokens
Optimized version of GPT-4.1 for speed and economy
Ideal for simple tasks on reasonable contexts
Total context: 128,000 tokens
Response length: up to 16,384 tokens
Ultra-lightweight version for simple tasks
Maximum efficiency for basic use cases
Total context: 128,000 tokens
Response length: up to 16,384 tokens
Balanced model between performance and efficiency
Versatile alternative to GPT-4.1
Total context: 128,000 tokens
Response length: up to 16,384 tokens
Lighter version of GPT-4o
Fast alternative for simple tasks
Total context: 128,000 tokens
Response length: up to 16,384 tokens
šÆ Selection Guide by Use Case#
1
For Versatile Use
GPT-4.1 offers the best performance/cost/versatility ratio with its reasoning capabilities and massive 1M tokens context, allowing processing of large documents while remaining less expensive than Gemini 2.5 Pro.
2
For Simple Tasks
GPT-4.1 Mini or Claude 3.5 Haiku constitute excellent economical choices on reasonable contexts.
3
For Massive Complex Document Analysis
Gemini 2.5 Pro excels in information extraction from voluminous content requiring reasoning, particularly when GPT-4.1 reaches its limits.
4
For Massive Contexts Without Complex Reasoning
Gemini 2.5 Flash optimizes costs for simple information extraction.
5
For Very Complex Reasoning
The o series models (o3, o3 Mini, o4 Mini) with their dedicated reasoning architecture, and GPT-5 with its unified system that automatically engages thinking mode for complex tasks.
6
For Reasoning on Medium Contexts
Claude Sonnet 4 offers excellent capabilities on intermediate data volumes.
7
For Unified Intelligence
GPT-5 offers a unique unified system that automatically switches between fast responses for simple queries and deep thinking for complex problems, eliminating the need to manually choose between different models.
8
For European Sovereignty
Pixtral Large represents the best French/European option, or Open Source models hosted at Scaleway.
The model choice directly impacts your application costs. More powerful models are generally more expensive but can prove more economical for complex tasks by requiring fewer iterations.š Cost Optimization#
š” Efficiency Strategies
Use "Mini" or "Haiku" models for simple and repetitive tasks
Prefer GPT-4.1 for its excellent quality/price ratio on complex tasks
Reserve Gemini 2.5 Pro for very large document analysis when necessary
Exploit open source models for less critical use cases
š Performance Monitoring
Use Thinkeo's integrated consumption tracking system to monitor costs per application and optimize your model choices according to your real needs.
āļø Parameters#
Parameters allow you to adjust more precisely the expected behavior from AI, which will influence the generated responses.š” Global Configuration: Parameters can be defined at the App level to apply to all AI blocks in the App. You can thus set up a global configuration for your AI blocks without having to change them one by one.
šļø Available Parameters#
Temperature
This parameter controls the degree of variability in model responses. A low temperature makes responses more deterministic and conservative, while a high temperature increases creativity and response diversity but can make them less coherent.Top P (or Nucleus Sampling)
This parameter limits token choices (words or word parts) based on their cumulative probability. If Top P is low, the model will choose among the most probable tokens. If Top P is high, this increases diversity of possible responses by including less probable tokens.Frequency Penalty
This parameter penalizes tokens that appear frequently in the generated response. A higher frequency penalty reduces word repetition, making the response more varied.Presence Penalty
This parameter penalizes tokens that already appear in the generated response, thus encouraging the model to introduce new concepts. A higher presence penalty increases diversity by discouraging the model from reusing the same words.Possible values will change according to the selected model. Temperature is dynamically adjusted when you switch from one model to another - remember to check that they correspond to your expectations.
š” Best Practice: It is not recommended to modify Temperature and Top P at the same time. Prefer one or the other in combination with one or two penalty parameters. We advise you to mainly work with Temperature, and possibly with frequency penalty.
š§ Reasoning Activation#
Certain AI models offer an advanced reasoning functionality that improves response quality for complex tasks requiring deep reflection. This capability transforms the traditional generation approach by allowing the model to "think" before responding.šÆ Models Compatible with Reasoning Activation in Thinkeo#
š¤ OpenAI - o Series (Dedicated Reasoning Models)
o3: Advanced model with deep reasoning
o3 Mini: Lightweight version with integrated reasoning
o4 Mini: Optimized model with reinforced reasoning capabilities
š Anthropic - Hybrid Models
Claude 4 Sonnet: Hybrid model with activatable reasoning
š Google - Models with Integrated Thinking
Gemini 2.5 Pro: Thinking model with automatic reasoning
Gemini 2.0 Flash Thinking: Experimental version dedicated to reasoning
ā” OpenAI - GPT-5 Series (Unified Intelligence)
GPT-5: Unified system with automatic routing between fast and thinking modes
GPT-5 Mini: Lightweight version with maintained reasoning capabilities
GPT-5 Nano: Ultra-efficient version for high-throughput applications
ā” How Reasoning Works#
When reasoning is activated, the model adopts a methodical approach that breaks down into several phases:1
Analysis Phase
The model breaks down the request into sub-problems and identifies key elements to address.
2
Reflection Phase
It explores different possible approaches and evaluates the implications of each choice.
3
Synthesis Phase
The model consolidates its reflection to produce a structured and reasoned response.
Problems requiring step-by-step analysis
Logical, mathematical, or scientific reasoning tasks
Complex questions requiring a methodical approach
Technical document analysis or multi-step problem solving
Cases where transparency of the thought process is important
š§ Configuration and Impact#
Reasoning activation is done directly in the AI block parameters. This option is only available if you have selected a compatible model.ā ļø Important Considerations:Processing Time: Reasoning significantly increases response time (from a few seconds to several minutes)
Cost: Reasoning models are generally more expensive to use
Quality: Response quality improvement often justifies the additional investment for complex tasks
šÆ Choosing the Right Reasoning Model#
For Balanced Reasoning: Claude Sonnet 4 offers an excellent compromise with its hybrid nature allowing switching between fast response and deep reasoning.For Very Complex Reasoning: OpenAI's o series models (o3, o3 Mini, o4 Mini) excel in dedicated reasoning tasks, while GPT-5 offers a unified system with automatic thinking mode activation for complex problems.For Massive Context Analysis: Gemini 2.5 Pro combines reasoning and ability to process very large volumes of data.
š§ System Prompts#
You can also add a complementary system prompt according to your prompting method. A system prompt can be used to specify a role or specific instructions that will facilitate AI interpretation or guide its understanding of the entire Prompt.Prompting is a subject in its own right on which you should iterate to achieve the desired result.Be clear, precise, and structured in your Prompts.
š Summary#
The AI Block is a powerful tool for creating dynamic, context-aware content generation within your Thinkeo applications. By understanding model capabilities, configuring parameters appropriately, and leveraging reasoning features when needed, you can create sophisticated AI-powered workflows that deliver high-quality, tailored outputs for your users.ā ļø An AI block must be executed from an Assistant/Wizard Step, either directly or via one of its parent blocks. It is always recommended to execute your App all at once.
Choose models based on your specific use case requirements
Configure parameters thoughtfully to achieve desired behavior
Utilize reasoning capabilities for complex analytical tasks
Structure your prompts clearly for optimal results
Monitor costs and performance to optimize your AI strategy
