Moving beyond Traditional RAG

Traditional RAG pipelines embed a query, retrieve context, and generate a response.

Agentic RAG introduces intelligence into the process. By classifying intent before deciding whether to retrieve, call tools, or answer directly, companies report cost reductions of up to 40% and latency improvements of 35%.

Core patterns driving adoption

Industry experts point to three architectural patterns that define Agentic RAG:

• Intent-Based Query Routing: determines whether retrieval is necessary or if a direct answer suffices.
• Tool Orchestration with Error Handling: coordinates APIs, calculators, and databases while managing failures gracefully.
• Continuous Cost & Latency Evaluation: tracks token usage and performance metrics in real time.

These patterns allow systems to decide, adapt, and optimise, a critical requirement for enterprise-scale AI.

Architecture in practice

Agentic RAG systems are typically built on three layers:

• Orchestration Layer: the “decision brain” that routes queries intelligently.
• Execution Layer: handles retrieval, tool calls, and LLM inference.
• Infrastructure Layer: provides vector databases, deployment management, and observability.

Unlike traditional RAG, which always retrieves, Agentic RAG evaluates whether retrieval is even necessary, orchestrating the optimal combination of retrieval, tools, and generation.

Provider flexibility through gateway layers

Another key trend is the rise of gateway abstractions that allow developers to switch seamlessly between providers such as OpenAI, Anthropic, Google, and Bedrock. This approach enables:

• Failover routing when providers experience downtime.
• A/B testing without code changes.
• Cost optimization by directing queries to the most efficient model.
• Freedom from vendor lock-in.

Companies are increasingly adopting unified gateways to balance speed, cost, and reliability across providers.

Conclusion

Agentic RAG is no longer a niche experiment but the blueprint for production AI systems. By combining retrieval with decision-making, orchestration, and observability, the technique is setting new standards for efficiency and adaptability in enterprise AI.

“Production AI isn’t about retrieval alone. It’s about intelligence: knowing when to retrieve, when to call tools, and when to answer directly. Agentic RAG delivers that intelligence”.

La entrada Agentic RAG for AI se publicó primero en Mosaic Factor.

Let’s break down the key types of AI and how we help you apply them.

AGI: the long-term horizon

General AI, or General Artificial Intelligence (AGI), refers to machines that can perform any intellectual task a human can. It’s flexible, autonomous, and capable of reasoning across domains.

Current Status: AGI is still theoretical. No existing system has achieved true general intelligence.

We think this is a highly compelling topic, and we are monitoring developments with great interest and anticipation.

Generative AI: real-world creativity at scale

Generative AI (or GenAI) is already transforming industries. These models create new content -text, images, code, audio, amongst others- based on learned data patterns.

Use Cases we can deliver for GenAI:

• Automated content generation for specific industries (like healthcare).
• Document summarisation and contract analysis for legal or compliance teams.
• Intelligent chatbots for internal company queries or customer support.
• Code generation and debugging tools for industry-specific developers (like automotive).

Our Solutions: we are capable of building and fine-tuning generative AI models using your proprietary data, ensuring outputs are accurate, brand-aligned, and compliant. Whether you need a custom GPT-style assistant or an image generator for product design, we can make it happen.

Specific AI techniques

Our core focus is enabling businesses to solve precise challenges through advanced AI techniques. It is what we have always done -what we call ‘traditional AI’ – and it has been central to our journey since our founding.

Our Approach: We design, train, and deploy these models with full lifecycle support: from data strategy and infrastructure to governance and performance monitoring.

Why partner with Mosaic Factor?

AI is powerful, but only when applied with precision. We don’t just deliver tools, we deliver transformation.

• Strategic AI consulting and roadmap development
• Custom model design and integration
• Scalable cloud and edge deployment
• Ongoing support, compliance, and optimisation

Whether you are exploring generative AI for creative automation or machine learning for operational efficiency, we help you turn potential into performance.

Ready to explore what AI can do for your business? Contact us to build something extraordinary, together.

La entrada General vs. Generative AI: what they mean for your business se publicó primero en Mosaic Factor.

Key Challenges for Automotive AI teams

The main challenges when developing AI models for SDVs are:

• High-Risk Classification: AI systems in SDVs often fall under the “high-risk” category, triggering mandatory conformity assessments, documentation, and audit readiness.
• Complex Compliance Landscape: with over 450 pages of legal text, translating regulation into engineering tasks is a daunting challenge. Manual reviews and fragmented documentation slow innovation and increase costs.
• Audit Pressure: teams must prepare for both scheduled and surprise inspections. Non-compliance can result in fines up to €7.5 million or 3% of global turnover.

AI Trust & Compliance frameworks

To address these hurdles, it is key to develop modular, automated compliance frameworks tailored for SDVs to enable:

• End-to-End System Assessments: rapid, audit-ready evaluations aligned with the EU AI Act.
• Configurable Reports: targeted assessments for specific AI components, such as dataset balance or model transparency.
• Audit Phase Interface: tools for third-party assessors, streamlining feedback and reducing evaluation time.

Who should worry about this?

• OEMs: responsible for system-level compliance, they must ensure every AI component meets regulatory standards before vehicle launch.
• Tier 1 Suppliers: developers of critical AI modules must demonstrate component-level compliance to OEMs, enhancing collaboration and market competitiveness.

Industry-wide implications

The EU AI Act is not just a legal hurdle, but a strategic opportunity. By embedding compliance into the development lifecycle, automakers can build more resilient, transparent, and future-proof AI systems.

The Act encourages:

• Cross-functional collaboration: AI, cybersecurity, safety, and regulatory teams must work in tandem.
• Lifecycle accountability: from design to post-market monitoring, traceability becomes a core requirement.
• Innovation through structure: automated tools and frameworks transform compliance from a bottleneck into a catalyst for better engineering practices.

As SDVs become the norm, the EU AI Act will shape how automotive AI is built, validated, and deployed. Forward-thinking developers and suppliers who embrace structured compliance will not only avoid penalties but will lead the next wave of intelligent mobility.

GPAI & GenAI Under the EU AI Act: what developers must know

The GPAI Code of Practice, finalised in May 2025, provides critical guidance for providers of General-Purpose AI (GPAI) models, including Generative AI (GenAI) systems. The EU AI Act distinguishes between complex GenAI models -those with systemic risk– and simpler GenAI models, each facing different compliance burdens:

• Complex GenAI (systemic risk models): These models exceed thresholds like 10²⁵ FLOPs in training compute or demonstrate high-impact capabilities across domains. Providers must conduct adversarial testing, risk assessments, and incident reporting, and ensure cybersecurity protections. They must also notify the European Commission for public database inclusion and maintain detailed documentation of model architecture and evaluation strategies.
• Simple GenAI Models: These are not classified as systemic risk and face lighter obligations. Providers must publish a summary of training data, ensure copyright compliance, and maintain technical documentation for downstream users. Transparency is key: outputs must be labeled, and users informed when interacting with AI systems.

The Code of Practice serves as a blueprint for demonstrating compliance, helping developers navigate the AI Act’s layered requirements while fostering innovation and trust in GenAI technologies.

La entrada Navigating the EU AI Act: what it means for SDV developers and Automotive innovators se publicó primero en Mosaic Factor.

Here are the key trends and developments that stood out.

Digital Twin Innovations

The fair featured numerous Digital Twin solutions, emphasizing virtual commissioning and real-time manufacturing services:

• Delta Electronics presented a Digital Twin for virtual commissioning of robotic gluing cells, enhancing automation efficiency.
• An open-source initiative by IDTA (Eclipse Foundation) garnered attention for advancing accessibility in Digital Twin technology.
• Circularity emerged as a critical theme, with contributions from Fraunhofer or Schneider promoting sustainable manufacturing practices. These projects spanned sectors such as food production and automotive component recycling.
• Siemens showcased a wind tunnel simulation using Digital Twin and AI technologies, it demonstrates the potential for advanced testing methods.

AI driving innovation

AI applications in manufacturing were prominently displayed, extending beyond traditional machine learning:

• The importance of Explainable AI (XAI) and Certified AI was highlighted, reflecting the industry’s commitment to transparency and reliability.
• Schneider utilised AI in plant-based milk production to optimise processes, ensuring consistent sugar content—an example of machine learning automating complex tasks.
• DFKI‘s AI solutions included temperature configuration analysis for heating systems and quality assurance for car body construction, combining image analysis, sensors, and Digital Twin technology.
• Robotics-integrated AI applications also introduced explainability for tasks such as detecting assembly errors and automating quality checks.

Robotics and AI convergence

The robotics exhibits presented diverse applications:

• Project ROX demonstrated process automation.
• Siemens unveiled projects combining robots with virtual PLCs, powered by AI and large language models (LLMs). One application involved assembling toys based on customer instructions via a digital interface.

Hannover Messe proved that the synergy of Digital Twins, AI, and Robotics is transforming manufacturing. These technologies not only enhance efficiency but also pave the way for sustainable and innovative production solutions.

La entrada Hannover Messe: Exploring the Future of Manufacturing se publicó primero en Mosaic Factor.

So we are talking about having open, compliant models based on diversity and ethics, at a European level.

Industry-specific LLMs

Working towards the development of industry-specific language models based on OpenEuroLLM models offers a unique opportunity for companies. These models not only democratise access to high-quality AI technologies but also allow precise customisation to meet the specific needs of each sector.

Key Benefits:

1. Adaptability and Precision: The models can be fine-tuned for specific applications, improving the accuracy and relevance of AI solutions in industrial contexts.
2. Regulatory Compliance: Developed within the European regulatory framework, these models ensure that AI solutions comply with current regulations, reducing legal and ethical risks.
3. Linguistic and Cultural Diversity: The multilingual capability of these models preserves linguistic and cultural diversity, enabling companies to operate effectively in multiple European markets.
4. Transparency and Community: The open nature of the project fosters collaboration and knowledge sharing, creating an active community of developers and users who can contribute to the continuous improvement of the models.

Based on Trustworthy AI

For companies, investing in the development of industry-specific language models based on OpenEuroLLM is not only an innovative strategy but also a way to ensure they are at the forefront of AI technology, complying with European standards and fully leveraging AI capabilities to enhance their competitiveness in the global market.

→ Check our LLMs solution

La entrada Open LLMs for AI transparency se publicó primero en Mosaic Factor.

To provide an idea, the top-5 algorithms we use more often for predictive models are:

1. Statistical Models: sophisticated statistical models and approaches such as generalised modeling, regularisation, Bayesian Inference, and time series analysis and forecasting which are used to capture intricate dependencies, model uncertainty, and make robust predictions with generalised models based on complex data distributions and latent structures.
2. Machine Learning Algorithms: powerful models to capture complex data relationships with tree-based, kernel-based, ensemble techniques (bagging, boosting, stacking and blending, and voting ensembles). The advanced supervised ML approaches are enhanced with techniques to improve generalisation and interpretability. Reinforcement learning through environment interactions by applying optimisation of policies, value-based learning, and actor-critic methods are designed for (sequential) decision-making.Advanced and tailored unsupervised learning techniques to focus on discovering hidden patterns, creation of segments and groups are developed and used. These techniques include:
   1. clustering,
   2. dimensionality reduction,
   3. and representation learning.
3. Deep Learning techniques: Deep Learning is based on deep neural networks to learn hierarchical representations of data, being key in applications such as natural language processing and image recognition.
4. Neural Networks are advanced models and approaches from deep learning. Representation learning, and attention-based architectures that enable state-of-the-art and also beyond-state-of-the-art with innovation in areas like computer vision, natural language processing, and sequential modelling. The motivation stands for:
   1. improving generalisation,
   2. scalability,
   3. and interpretability through advanced techniques by pushing the boundaries of what a machine can learn.
5. Explainable Artificial Intelligence (XAI techniques): methods aiming to uncover how models with complex dataset and structure make the predictions, providing transparency in decision-making pipelines and processes. Techniques include both model-agnostic and model-specific approaches; they are crucial to understand the rationale behind a model output and a decision.

→ Check our Predictive Model solutions as well as our Trustworthy AI solutions.

La entrada Our top algorithms for predictive modeling se publicó primero en Mosaic Factor.

What are LLMs?

LLMs (Large Language Models), are advanced AI systems capable of understanding and generating various forms of content, including text, code, images, video, and audio. These models are trained on at least one billion parameters (data points), which allow them to grasp language patterns and respond appropriately.

LLMs find applications in natural language processing tasks such as text generation, translation, sentiment analysis, data analysis, question answering, and text summarisation.

Evolution of LLMs

Key milestones include:

• 1966 ELIZA: The first chatbot simulating a psychotherapist.
• 2013 word2vec: Efficient methods for learning word embeddings from raw text.
• 2018 GPT and BERT: Groundbreaking models.
• 2020 GPT-3: A significant leap.
• Late 2021 and 2022: GPT-4 and other advancements.
• Statistical models: Developed to learn patterns from text data.

LLMs vs. NLP

While NLP (Natural Language Processing) models interpret or transform existing text, LLMs excel at generating new, coherent text from scratch.

They can create essays, stories, and even computer code that mimics human writing styles.

Light LLMs

Nowadays, though, there is an increasing importance of smaller models (light LLMs) for specific domain applications.

While the largest models would all be “general purpose”, light LLMs are developed with a specific sector use in mind.

That is:

• Large models use a huge number of parameters, without tuning to a specific use, use a lot of energy, sometimes with questionable reliability, and that provide answers even when they don’t know them.
• Smaller models consider the use that is going to be given to it, refining its responses (fine-tuning) the specific model for a specific use.

Light LLMs benefits

1. Efficiency: Light LLMs require fewer computational resources, making them faster and more cost-effective.
2. Scalability: Companies can deploy light LLMs across various applications without straining infrastructure.
3. Customisation: Light models allow fine-tuning for specific tasks, tailoring them to company needs.
4. Privacy: Smaller models reduce the risk of inadvertently leaking sensitive information.
5. Easier Maintenance: Light LLMs are simpler to manage and update.

To conclude, while both open-source and closed LLMs have their merits, light LLMs offer practical advantages for companies seeking efficient, adaptable solutions. Therefore, you should consider your specific requirements when choosing the right LLM for your organisation.

→ Check our LLMs solution

La entrada What are light LLMs? se publicó primero en Mosaic Factor.

The explainability of an AI model can be put in practice in different ways for various industries. Let’s see some examples.

1.Healthcare

When working in healthcare, we are talking about highly regulated environments that need to be certified, trusted and accountable. For instance, when performing patient disease diagnosis, explainable AI can explain the elements and data that was used to diagnose that patient. This way, we help create greater trust between patients and their doctors while mitigating any potential ethical issues when a machine is aiding the detection of a disease.

Typical use cases for this are validation of AI predictions that work with medical imaging data when diagnosing cancer.

2.Manufacturing

Explainable AI can also by applied in a production line to detect, map and explain the causes for unproper machine behaviour or defective product outputs, causing what it is called “nonconformities” on product quality in the production process or highlight the need for maintenance.

This way there is a higher understanding of machine-machine and machine-operator communication and business management policies can be made to decrease costs and gain productivity while keeping trusted and save production standards that need to be complaint and certified.

3.Mobility

Explainable AI is becoming increasingly important in the transport and automotive industry due to the expansion of IoT and smart mobility solutions as well as the potential expansion in use of autonomous vehicles -first in business environments such as logistics self-driven vehicles or trains, later in end-users.

This has placed an emphasis on explainability techniques for AI algorithms, especially when it comes to using cases that involve safety-critical decisions. Explainable AI can be used for autonomous vehicles where it provides increased situational awareness in accidents or unexpected situations, which could lead to more responsible technology operation (i.e., preventing crashes).

4.Recruitment

Resume screening: explainable artificial intelligence could be used to explain why a resume was selected or not. This provides an increased level of understanding between humans and machines, which helps create greater trust in AI systems while mitigating issues related to bias and unfairness.

5.Finance

Fraud detection: Explainable AI is important for fraud detection in financial services. This can be used to explain why a transaction was flagged as suspicious or legitimate, which helps mitigate potential ethical challenges associated with unfair bias and discrimination issues when it comes to identifying fraudulent transactions.

Loan approvals: explainable artificial intelligence can be used to explain why a loan was approved or denied. This is important because it helps mitigate any potential ethical challenges by providing an increased level of understanding between humans and machines, which will help create greater trust in AI systems.

→ Check our Mosaic XAI dashboards

La entrada Trustworthy AI for different industries se publicó primero en Mosaic Factor.

1. Unacceptable risk is prohibited. Therefore, the following types of models should not be used:
   1. a. Subliminal, manipulative, or deceptive AI
   2. b. Systems that exploit vulnerabilities related to age, disability, or socio-economic circumstances to distort behaviour, causing significant harm.
   3. c. Biometric categorisation systems inferring sensitive attributes (race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation), except labelling or filtering of lawfully acquired biometric datasets or when law enforcement categorises biometric data.
   4. d. Social scoring systems.
   5. e. Systems assessing risk of an individual committing criminal offenses.
   6. f. Compiling facial recognition databases from the internet or CCTV footage.
   7. g. Inferring emotions in workplaces or educational institutions.
   8. h. ‘Real-time’ remote biometric identification (RBI) in publicly accessible spaces for law enforcement.
2. High-risk AI systems: they are regulated, and the AI Act focuses mostly on these.
3. Limited risk AI systems: they are subject to lighter transparency obligations. This means, developers and deployers should make sure that end-users are aware that they are interacting with AI (thus making it clear that there is an AI model behind chatbots and deepfakes).
4. Minimal risk AI models are unregulated: these include most AI applications that were available on the EU single market at the moment of starting AI Act in 2021. For instance, AI-enabled video games and spam filters.

Clearly, this scenario is changing with generative AI, which raises the level of risk of AI models, making them mostly high-risk. Even if now AI regulations are advancing, we think AI systems should be overseen by people, rather than by automation, to prevent harmful outcomes. This includes making sure we are able to generate trustworthy AI systems that are fair by design and are explainable and clear to decision makers. Legislation will clearly not fast-track adoption of responsible AI but organisations are the ones who need to share experiences and solutions to show what “good practice” looks like. Boards need to embrace Corporate Digital Responsibility to assess digital impacts of products/services on all stakeholders by examining societal, economic, technological & environmental impacts. Therefore, we are supporting companies in their role to ensure the technology is not deployed in “negative use cases” that could harm society and generating AI models that are transparent and accountable across industries.

→ Check our Trustworthy AI solution

La entrada Key aspects of the european AI act se publicó primero en Mosaic Factor.