Compliance architecture
Is your RAG pipeline GDPR compliant?
The moment personal data is embedded into a vector index, retrieval becomes processing under GDPR. Data protection by design does not stop at the source system: if the copilot can hand anyone's HR file to any employee who asks a similar-enough question, Articles 5, 25, and 32 have something to say about your architecture.
Gateco status, stated plainly
Gateco can be deployed in a GDPR-compliant configuration and a Data Processing Agreement is available. Compliance remains a property of your deployment and lawful basis, not a vendor badge. Details on the Gateco GDPR page.
Where personal data leaks in a standard RAG pipeline
Source systems enforce access: the HR system knows who may read a personnel file, the DMS knows who may open a contract. Embedding copies the content out from under those controls. The vector database ranks by similarity, the service account can read everything, and no record exists that would satisfy an Article 30 request. A data subject access request or a regulator's question lands on a pipeline that cannot answer who saw what.
GDPR requirement to retrieval control
| GDPR requirement | How standard RAG fails it | Retrieval-layer control |
|---|---|---|
| Integrity and confidentiality (Art. 5(1)(f)) | Any authenticated employee can retrieve any embedded personal data via similarity search. | Per-chunk, per-user policy evaluation before content reaches the model; deny-by-default. |
| Data protection by design (Art. 25) | Authorization was designed for the source systems and silently dropped at embedding time. | The enforcement point sits in the retrieval path itself, so protection survives the data's change of shape. |
| Security of processing (Art. 32) | One service credential with full index read access is the entire security model. | Named principals from your IDP, classification ceilings, envelope-encrypted connector credentials. |
| Records of processing (Art. 30) | No demonstrable record of which personal data the AI accessed, for whom, or under what rule. | Every retrieval decision is logged with principal, resource, policy, and verdict; exportable for the DPO. |
| Right to erasure (Art. 17) | Deleting the source document leaves its chunks retrievable from the vector index. | Resource deletion removes vectors, chunk registry, and the gated resource together; source syncs propagate tombstones. |
When you do not need a retrieval authorization layer
If the corpus holds no personal data, GDPR's retrieval questions largely fall away. If it holds personal data but every user is equally entitled to it under one purpose, your remaining obligations are records, security of storage, and erasure, and a full policy layer may be more than you need. The architecture on this page earns its place when personal data with different audiences shares an index, which is what an internal copilot over HR, legal, and operations documents looks like.