Training
Operational control and regulatory compliance after deployment
This course addresses operational compliance for Retrieval-Augmented Generation (RAG) systems under Regulation (EU) 2024/1689 (the EU AI Act). A RAG system generates output by retrieving material from an external knowledge source and incorporating that material into the response.
Whether a RAG system is high-risk depends on intended purpose, not on the architecture. Where that purpose places the system in a high-risk class, the Act continues after the system is put into service. Providers and deployers then carry duties on risk management, data and data governance, logging, human oversight, accuracy and robustness, quality management, use in operation, and post-market monitoring.
The course translates the articles named for this programme into purpose-specific controls for a live RAG pipeline: what is retrieved, from where, how it is ranked, how it is used in the generated output, and how that chain is watched after go-live.
A RAG system has two moving parts the Act will see as one system in use: the retriever and the generator. Errors can sit in either part, or in the join. A biased, stale or incomplete corpus produces a different failure from a model that invents a citation. Source drift - the corpus changing after the system was tested - is a production fact, not a design footnote.
Those facts change how the listed articles are applied. Data governance has to reach the corpus and the retrieval index, not only the training set. Oversight has to see retrieved passages and generated text. Logs have to reconstruct the query, the hits, the sources used and the output. Monitoring has to cover retrieval quality, hallucination and the propagation of bias from sources into answers.
The course is built around the articles already named for this programme.
The course does not treat every RAG deployment as high-risk. It does not treat high-risk status as optional once intended purpose meets the Act's tests.
By the end of the course, participants will be able to:
Six modules. Each module stays inside the articles listed above.
How the Act classifies systems by intended purpose. When a RAG use is high-risk and when it is not. How intended purpose, state of the art and the risk-management system are read together so compliance is attached to the live pipeline.
What a person must be able to see: the query, the retrieved material, the sources, and the generated answer. When oversight must stop or correct an output. Escalation when retrieval is empty, conflicting or clearly wrong, and when the generator departs from the retrieved material.
Article 12 as it applies to a retrieval-plus-generation sequence. Reconstructing which sources were hit and which were used. Article 10 as it applies to the corpus and the index. Article 15 as it applies to accuracy and robustness of the joined system.
Collecting and analysing data on how the RAG system performs in use. Retrieval quality over time. Source drift after the last test. Hallucination rates that matter for the intended purpose. Bias that moves from the corpus into answers. The deployer's related duty under Article 26 to use the system as instructed and to watch its operation.
Article 9 after deployment: risks that appear only when the corpus, the index and the generator are connected. Containing a bad source. Stopping a pipeline that is retrieving from an unintended store. Recording the incident so the risk-management file can be updated.
How Articles 17 and 8 require these operational controls to sit inside a quality-management system. What a notified body or market-surveillance authority would need in order to assess the live RAG system against the claimed purpose. The record means the technical and operational information the Act already requires. The course does not add a separate product pack.
The course is written for people responsible for RAG systems that are in production, or that are being prepared for a high-risk intended purpose under the articles above. It assumes familiarity with the EU AI Act. It is not a survey of retrieval methods.