News
News
News
Struck AI Usage
1 mei 2025
Gecreëerd door
Struck Team
Communication
Heb je een vraag?
Neem contact op met ons team als je vragen hebt
Neem contact op met ons team als je vragen hebt
Struck AI Usage
Struck utilises a modular approach to AI, selecting the best model for a given application based on its unique strengths. This enables more tailored, efficient, and powerful responses compared to generalist systems like ChatGPT or Microsoft Copilot.
Why Multiple Models?
Different AI models offer different capabilities:
Multimodal capabilities (e.g. processing images and tables) are only supported by some.
Context window limitations can affect how much text a model can handle.
Performance trade-offs exist — for example, O1 excels in complex reasoning but is slower and more limited in availability.
Gemini offers both “Flash” and “Pro” editions, where Flash is faster but less nuanced, and Pro is better at handling complex or large-context tasks.
Analogy: Think of Struck’s system as a team of specialists, each chosen based on the task, whereas generalist AI services operate like an army of one-size-fits-all responders.
Agentic AI at Struck
Struck uses an agentic AI approach, where tasks are defined as goals, and the system is given decision-making ability to pursue them.
Example: At Loeman’s, a tender evaluation system uses a set of standardised questions to generate a detailed report on the attractiveness of a project.
The Struck Pipeline: Simplified Question-to-Response Flow
Language Detection
Detects the user’s input language and target output language.Query Analysis
Interprets the user’s question and extracts context clues from the conversation history.Retrieval
Searches both internal regulatory databases and user-provided documents.Reranking & Relevance Filtering
Applies stricter criteria to filter and rank retrieved content.Chat Completion
Generates a coherent and contextually accurate response using all gathered data.
Model Usage Table
Usage | Model | Description |
|---|---|---|
Chat completion | gpt-4o | This model takes all of the gathered sources and generates the answer response. |
Zoning Chat Completion | gemini-2.0-pro-001 | Same as above, but benefits from larger context window. |
Deep Analysis | o1 | Complex reasoning and in-depth chain of thought reasoning. |
Retrieval Evaluation | Custom trained models, with training data from country specific building codes | Commonly known as reranking is a process, in which after a set of documents is identified as relating to a query, the “reranker” will apply a more thorough process of vetting the content to ensure that 6he best possible text is selected to be used by later steps. |
Query Analysis rephrasing/classification | gpt-4o-mini | Review the user’s question and rephrase it to take previous chat history into consideration. |
Query metadata extraction | gemini-2.0-flash-001 | Extract facts, figures, or context clues. For example: “I'm renovating a building in Nijmegen …" We can extract the relevant region and scope our regulations queries to include regional documents for Nijmegen. |
Language Detection | gpt-4o | Detect the user’s input language. |
Title Generation | gemini-1.5-flash | Generates a title to summarise the chat description for the UI presentation. |
hyde | gpt-4o | Hypothetical Document Embeddings. This rephrases the question in the native language of the source documents and generates related terms to improve search and document retrieval. |
Keyword Processing | Gemini 2.0 Flash | Extract relevant terms and use industry relevant jargon instead of what the user typed when it will improve results. |
Document Summary | gemini-1.5-pro-001 | Summarise document and identify key takeaways |
Document Ingestion | gpt-4o | |
Image Analysis | gpt-4o, gemini-2.0-flash | Both are used mainly for quota reasoning |
Monument Status | gemini-2.0-flash | Searching archives like monument.nl and others and extracting a response for a target location. |
Translation | gemini-1.5-pro-001 | When a user requests that a summary be translated into another language |
Privacy Considerations
Struck prioritises data privacy by using models hosted in Europe, avoiding exposure to U.S. data regulations. All model providers have agreements that prohibit training on user data.
Struck AI Usage
Struck utilises a modular approach to AI, selecting the best model for a given application based on its unique strengths. This enables more tailored, efficient, and powerful responses compared to generalist systems like ChatGPT or Microsoft Copilot.
Why Multiple Models?
Different AI models offer different capabilities:
Multimodal capabilities (e.g. processing images and tables) are only supported by some.
Context window limitations can affect how much text a model can handle.
Performance trade-offs exist — for example, O1 excels in complex reasoning but is slower and more limited in availability.
Gemini offers both “Flash” and “Pro” editions, where Flash is faster but less nuanced, and Pro is better at handling complex or large-context tasks.
Analogy: Think of Struck’s system as a team of specialists, each chosen based on the task, whereas generalist AI services operate like an army of one-size-fits-all responders.
Agentic AI at Struck
Struck uses an agentic AI approach, where tasks are defined as goals, and the system is given decision-making ability to pursue them.
Example: At Loeman’s, a tender evaluation system uses a set of standardised questions to generate a detailed report on the attractiveness of a project.
The Struck Pipeline: Simplified Question-to-Response Flow
Language Detection
Detects the user’s input language and target output language.Query Analysis
Interprets the user’s question and extracts context clues from the conversation history.Retrieval
Searches both internal regulatory databases and user-provided documents.Reranking & Relevance Filtering
Applies stricter criteria to filter and rank retrieved content.Chat Completion
Generates a coherent and contextually accurate response using all gathered data.
Model Usage Table
Usage | Model | Description |
|---|---|---|
Chat completion | gpt-4o | This model takes all of the gathered sources and generates the answer response. |
Zoning Chat Completion | gemini-2.0-pro-001 | Same as above, but benefits from larger context window. |
Deep Analysis | o1 | Complex reasoning and in-depth chain of thought reasoning. |
Retrieval Evaluation | Custom trained models, with training data from country specific building codes | Commonly known as reranking is a process, in which after a set of documents is identified as relating to a query, the “reranker” will apply a more thorough process of vetting the content to ensure that 6he best possible text is selected to be used by later steps. |
Query Analysis rephrasing/classification | gpt-4o-mini | Review the user’s question and rephrase it to take previous chat history into consideration. |
Query metadata extraction | gemini-2.0-flash-001 | Extract facts, figures, or context clues. For example: “I'm renovating a building in Nijmegen …" We can extract the relevant region and scope our regulations queries to include regional documents for Nijmegen. |
Language Detection | gpt-4o | Detect the user’s input language. |
Title Generation | gemini-1.5-flash | Generates a title to summarise the chat description for the UI presentation. |
hyde | gpt-4o | Hypothetical Document Embeddings. This rephrases the question in the native language of the source documents and generates related terms to improve search and document retrieval. |
Keyword Processing | Gemini 2.0 Flash | Extract relevant terms and use industry relevant jargon instead of what the user typed when it will improve results. |
Document Summary | gemini-1.5-pro-001 | Summarise document and identify key takeaways |
Document Ingestion | gpt-4o | |
Image Analysis | gpt-4o, gemini-2.0-flash | Both are used mainly for quota reasoning |
Monument Status | gemini-2.0-flash | Searching archives like monument.nl and others and extracting a response for a target location. |
Translation | gemini-1.5-pro-001 | When a user requests that a summary be translated into another language |
Privacy Considerations
Struck prioritises data privacy by using models hosted in Europe, avoiding exposure to U.S. data regulations. All model providers have agreements that prohibit training on user data.
Struck AI Usage
Struck utilises a modular approach to AI, selecting the best model for a given application based on its unique strengths. This enables more tailored, efficient, and powerful responses compared to generalist systems like ChatGPT or Microsoft Copilot.
Why Multiple Models?
Different AI models offer different capabilities:
Multimodal capabilities (e.g. processing images and tables) are only supported by some.
Context window limitations can affect how much text a model can handle.
Performance trade-offs exist — for example, O1 excels in complex reasoning but is slower and more limited in availability.
Gemini offers both “Flash” and “Pro” editions, where Flash is faster but less nuanced, and Pro is better at handling complex or large-context tasks.
Analogy: Think of Struck’s system as a team of specialists, each chosen based on the task, whereas generalist AI services operate like an army of one-size-fits-all responders.
Agentic AI at Struck
Struck uses an agentic AI approach, where tasks are defined as goals, and the system is given decision-making ability to pursue them.
Example: At Loeman’s, a tender evaluation system uses a set of standardised questions to generate a detailed report on the attractiveness of a project.
The Struck Pipeline: Simplified Question-to-Response Flow
Language Detection
Detects the user’s input language and target output language.Query Analysis
Interprets the user’s question and extracts context clues from the conversation history.Retrieval
Searches both internal regulatory databases and user-provided documents.Reranking & Relevance Filtering
Applies stricter criteria to filter and rank retrieved content.Chat Completion
Generates a coherent and contextually accurate response using all gathered data.
Model Usage Table
Usage | Model | Description |
|---|---|---|
Chat completion | gpt-4o | This model takes all of the gathered sources and generates the answer response. |
Zoning Chat Completion | gemini-2.0-pro-001 | Same as above, but benefits from larger context window. |
Deep Analysis | o1 | Complex reasoning and in-depth chain of thought reasoning. |
Retrieval Evaluation | Custom trained models, with training data from country specific building codes | Commonly known as reranking is a process, in which after a set of documents is identified as relating to a query, the “reranker” will apply a more thorough process of vetting the content to ensure that 6he best possible text is selected to be used by later steps. |
Query Analysis rephrasing/classification | gpt-4o-mini | Review the user’s question and rephrase it to take previous chat history into consideration. |
Query metadata extraction | gemini-2.0-flash-001 | Extract facts, figures, or context clues. For example: “I'm renovating a building in Nijmegen …" We can extract the relevant region and scope our regulations queries to include regional documents for Nijmegen. |
Language Detection | gpt-4o | Detect the user’s input language. |
Title Generation | gemini-1.5-flash | Generates a title to summarise the chat description for the UI presentation. |
hyde | gpt-4o | Hypothetical Document Embeddings. This rephrases the question in the native language of the source documents and generates related terms to improve search and document retrieval. |
Keyword Processing | Gemini 2.0 Flash | Extract relevant terms and use industry relevant jargon instead of what the user typed when it will improve results. |
Document Summary | gemini-1.5-pro-001 | Summarise document and identify key takeaways |
Document Ingestion | gpt-4o | |
Image Analysis | gpt-4o, gemini-2.0-flash | Both are used mainly for quota reasoning |
Monument Status | gemini-2.0-flash | Searching archives like monument.nl and others and extracting a response for a target location. |
Translation | gemini-1.5-pro-001 | When a user requests that a summary be translated into another language |
Privacy Considerations
Struck prioritises data privacy by using models hosted in Europe, avoiding exposure to U.S. data regulations. All model providers have agreements that prohibit training on user data.
