AI does not make final technical decisions.

Vectoria Civil can assist with explanations, report preparation, checks, summaries and guided use of tools. The user decides, reviews and validates every result.

What the agent can do

The agent can interpret user instructions, read project context supplied to it, propose steps, run available tools and draft responses or reports with traceability where the workflow supports it.

Data that may be sent

Conversation text, technical summaries, tool results, geometry parameters, project excerpts and files or documents that the user chooses to attach or use in a specific feature may be processed.

The remote agent processes that information through Vectoria Civil infrastructure and TensorX Limited, an inference processor established in Ireland. TensorX runs open models from the DeepSeek family in Dublin and Helsinki without sending prompts or completions to the model developer; it states that inference content is retained for zero time and is not used for training. The backend blocks the agent if the key, recorded DPA acceptance or approved configuration is missing, and does not redirect content to Gemini, Vertex AI, OpenAI or another fallback provider. If the user selects Codex, the application uses the locally installed Codex client and the OpenAI account signed in by the user; submitted instructions, limited history, and context are processed under the terms and privacy notice applicable to that account. Roles, retention and transfers are explained in the Privacy Notice.

Voice

Voice is optional and is captured only while the microphone is active. Audio is streamed through the Vectoria Civil backend to Google Cloud Speech for transcription. The feature code does not deliberately retain audio after the session. The transcript first appears in the instruction field, and the user decides whether to submit it to the agent.

Data you should not send

Do not submit unnecessary personal data, tax information, secrets, credentials, signed URLs, confidential contracts, third-party data without a lawful basis, or technical information that is not necessary for the task. The beta prohibits GDPR Article 9 special-category data, criminal data, identification biometrics, medical or health information, children's data and full payment-card numbers. There is no standard authorisation to process them. Professional customers who add ordinary third-party personal data must comply with the customer–Vectoria DPA.

Known limitations

AI may be wrong, omit information, misunderstand an instruction, produce incomplete text or depend on insufficient inputs. It does not replace applicable regulations, professional judgement, technical supervision, independent validation or the user's professional responsibility.

How to review results

Privacy and logs

Visible conversation history is kept in a local database until the user deletes it. Vectoria Civil creates local agent logs containing prompts, actions, arguments, result previews, activities, and messages; Codex creates separate local logs with a limited prompt, recent history, summarized context, response, and tool events. Both apply size limits and redaction of known secret patterns; the application deletes them on the next application start or log write after they exceed 30 days. New logs linked exactly to a conversation are also deleted when that conversation is deleted; legacy logs without an identifier expire through the same 30-day cleanup cycle or through an explicit global local purge. Redaction is not infallible: do not submit secrets or unnecessary data.

The backend does not retain prompt text or documents as an AI case record, but keeps usage events for no more than 30 days with identifiers, task, model, token, cost, credit, status, and operational metadata, without prompt text or file content. Cloud Logging keeps aggregate metrics, statuses, and error classes for no more than 30 days without raw prompt, tool-argument, or project values.

Fail-closed controls

Remote features require an authenticated account, license, and enabled configuration. If a required provider, credential, permission, or gate is missing, the feature is blocked and reported as unavailable; content is not sent to an undeclared fallback provider.