Context Library security — how your uploaded documents are protected

This article explains what happens to a document you upload to the Context Library: where it is stored, who can see it, how the Advisor works with it, and how you delete it again. It is written so you can pass the answers straight to your data protection officer or your IT security team.

In short

1. Your data stays in the EU. Storage and processing run in the AWS Frankfurt region (`eu-central-1`).
2. Your documents stay inside your organisation. Every search is technically bound to your tenant. Other customers never see your content.
3. You decide who reads along. Uploads to "My Context" are visible to you alone. "Company Context" is released to your organisation by an administrator.
4. Your content does not train any AI model. Neither your documents nor the Advisor's answers feed into the training of foundation models — guaranteed both contractually and technically.
5. You can delete completely, at any time. One action removes the original file, the extracted text and the search index together.

What happens when you upload a document

A file travels through five steps:

1. Transfer. The file is transmitted over an encrypted connection (TLS 1.3 or higher). Your credentials are verified before anything is stored.
2. Storing the original. The original file is placed in object storage. Its storage path always begins with your organisation's identifier and is built server-side from your verified credentials — never from values supplied with the upload.
3. Text extraction. The text is read out of the file inside our own system. There is no external conversion service that your file is handed to.
4. Splitting into sections. The text is divided along its headings, so the Advisor can later find individual passages rather than whole documents.
5. Building the search index. Each section gets a mathematical representation that makes similarity search possible. This step also runs in the AWS Frankfurt region.

Supported formats are PDF, DOCX, HTML, TXT and MD. The limits are 20 MB per file and, by default, 100 MB of total storage per organisation. While a document is still being processed the Advisor cannot find it — only a fully indexed document is ever searched.

Where your documents are stored

Every component sits in the European Union, in the AWS Frankfurt region (`eu-central-1`):

Physical security of the data centres is AWS's responsibility under ISO 27001 and SOC 1/2/3. For transport and storage, the measures committed in our data processing agreement apply: TLS 1.3 or higher in transit, AES-256 at rest, using Customer Managed Keys held in a FIPS 140-2 HSM. Backups are taken daily, spread across multiple availability zones, and roll off automatically after 30 days.

Communication with the AI models runs exclusively over private VPC endpoints inside the AWS network. Your content never leaves that network over the public internet.

Who can see your documents

The Context Library has three areas. The area you choose at upload time determines visibility:

This rule is not merely display logic in the interface. It is part of every database query: each search is bound to your organisation, and documents in "My Context" are additionally bound to the identifier of the person who owns them. A query without that binding does not exist in the system.

Inside talentsconnect, need-to-know applies with role-based permissions. All administrative access is protected by multi-factor authentication, and infrastructure is reachable only over VPN or the AWS Session Manager.

How the Advisor finds and uses your documents

The Advisor does not read along in your documents continuously. It searches deliberately, when your question calls for it:

1. It knows the names, not the contents. The Advisor receives a list of the available documents — name, area and tags, but no content. That lets it judge whether searching is worthwhile at all.
2. It searches only when needed. The search is a deliberate step the Advisor takes when your question requires knowledge about your organisation. For a general question, your documents are never touched.
3. It searches two ways at once. Once by meaning, so that related wording is still found, and once by exact wording, so that names, key figures and abbreviations are found reliably. The two result lists are then merged into a single ranking. This also works when you ask in English and the document is written in German.
4. It receives passages, not whole documents. Typically the five best sections come back. Hits that score far below the best hit are filtered out beforehand.
5. It names its source. Answers built on your documents open with a reference to the document they came from. If the Advisor finds nothing, it says so — it does not invent a source.

One point matters especially to your IT security team: the content of your documents is deliberately treated as quoted material, never as instructions. Text inside a document therefore cannot issue new commands to the Advisor. In addition, we screen every document at upload for patterns typical of such manipulation attempts and flag conspicuous documents in the interface.

Attaching documents to a specific question

The "+" control in the chat attaches documents directly to a single question. It is the fastest route when you already know what you want to talk about.

• You can attach up to 20 documents to one question.
• Attached documents are searched with priority and occupy the majority of the available slots in the result list.
• One slot always stays free for the regular search. That way a genuinely relevant hit from the rest of your library is not lost if an attached document turns out not to fit the question.
• Attaching does not widen your permissions. At search time your visibility is checked again — the check made in the interface is enforced a second time in the database.

Tips for good results

• Meaningful file names. The Advisor sees your document names before it searches. "HR strategy 2026" leads to better decisions than "document_final_v3".
• Use real headings. Documents are split into sections along their headings. A well-structured document produces more precise answers.
• Text PDFs rather than scans. A scanned PDF without a text layer contains no readable text and the upload will fail. No text recognition (OCR) takes place.
• Delete what is outdated. The Advisor does not judge which document is the more current one. If two versions sit in the library, both can appear in an answer.
• Same name replaces. Upload a document under the same name again and the new version fully replaces the old one.

What we do not do with your content

• No model training. There is a technical and contractual guarantee that neither your inputs nor the generated answers are used to train or improve foundation models — not by AWS, and not by any third party.
• No persistent storage at the model provider. Input data is held in volatile memory for the duration of processing and is not retained afterwards.
• No access for model providers. Data flows are logically isolated from other AWS customers and from the model providers. Providers of the models have no access to inputs or outputs.
• No public internet. The connection to the models runs over private network endpoints inside AWS.
• No sharing with other customers. There is no shared knowledge store across tenants. Your documents ground answers inside your organisation only.
• No external document service. Reading PDF, DOCX and HTML happens inside our own system.

Deleting documents

You delete a document directly in the Context Library. Deletion is final and covers, in one step:

• the original file in object storage,
• the extracted text,
• all sections and their search index.

There is no recycle bin and no remaining copy marked as deleted. Immediately after deletion the Advisor can no longer find or cite the document. Database backups expire on their normal 30-day cycle.

Who may delete follows the same rule as uploading: you delete your own documents in "My Context", while Company Context is managed by the administrators in your organisation.

Limits we state openly

Trust also comes from saying what the system does not do:

• No text recognition. Scanned PDFs without a text layer cannot be processed.
• Size limits. 20 MB per file, 100 MB per organisation by default.
• Format limits. PDF, DOCX, HTML, TXT and MD only.
• There are no automatic deletion periods. Documents remain in the library until someone deletes them. Control over retention deliberately stays with you.
• No candidate documents. Do not upload CVs, application documents or special categories of personal data under Art. 9 GDPR. The Advisor is not a candidate assessment tool — use Talent Studio for that.

Questions your data protection officer will ask

Where is the data processed and stored?

Exclusively in the European Union, in the AWS Frankfurt region (`eu-central-1`). No transfer to third countries takes place.

Is our content used to train AI models?

No. There is a technical and contractual guarantee that neither inputs nor outputs are used to train foundation models. The model provider has no access to your content.

Who has access to our documents?

Within your organisation, the people the chosen area provides for — with "My Context", only the person who uploaded the document. On the talentsconnect side, need-to-know applies with role-based permissions and multi-factor authentication on all administrative access.

Can we have data deleted completely?

Yes, and you do it yourself. Deletion removes the original file, the text and the search index together and cannot be undone. Backups expire after 30 days.

Is there a data processing agreement?

Yes. talentsconnect provides a DPA under Art. 28 GDPR including technical and organisational measures and a list of sub-processors. Your contact is datenschutz@talentsconnect.com.

How long is data retained?

Uploaded documents remain stored until you delete them — there is no automatic deletion period. Separate retention periods apply to conversation data.

Questions your IT security officer will ask

How is tenant separation implemented?

Every query against the document stores is bound to the organisation identifier taken from the authentication token, and personal documents are additionally bound to the user identifier. Access runs through exactly one encapsulated code path, which keeps this rule auditable. Storage paths in object storage likewise begin with the organisation identifier, so a cross-tenant path cannot be constructed.

How is data encrypted?

In transit with TLS 1.3 or higher. At rest with AES-256 using Customer Managed Keys held in a FIPS 140-2 HSM, which means cloud provider staff have no access to the key material.

How are authentication and authorisation handled?

Through Keycloak with OIDC. Every request carries a verified token, and organisation membership is validated against the issuing realm. Permission to write to Company Context is tied to administrator roles.

What if someone hides instructions inside a document?

Document content is passed to the Advisor explicitly as quoted material, not as instructions — that is the effective control. In addition, we screen every document at upload for typical manipulation patterns and mark conspicuous documents visibly in the interface.

Which sub-processors are involved?

For processing uploaded documents this is AWS (storage and AI processing via Amazon Bedrock in Frankfurt). The complete, contractually agreed list of sub-processors is part of the DPA.

Are the data centres certified?

The AWS data centres in use are ISO 27001 certified and audited under SOC 1/2/3. We provide the certificates on request.

Further documentation

• Data processing agreement (DPA) with its annex on technical and organisational measures and the list of sub-processors — on request or through your talentsconnect contact.
• Instructions for Use of the AI Advisor — intended purpose, limitations and rules for human oversight.
• Questions, evidence and incident reporting: datenschutz@talentsconnect.com.