Keep projects scoped
Sources, reviewed corrections, retrieval indexes, and adapter versions are scoped to their owner and project. Owners explicitly grant eligible workspace members project access; membership alone does not share raw source files.
Security & data controls
Choose the material, define who can work with it, and review how it shapes your model. Clear access and processing boundaries support responsible organizational AI.
01 / Your institutional knowledge
Sources, reviewed corrections, retrieval indexes, and adapter versions are scoped to their owner and project. Owners explicitly grant eligible workspace members project access; membership alone does not share raw source files.
Training produces a versioned candidate. Artifact validation, evaluation, and matching runtime evidence are checked before activation, keeping the release decision tied to the model that was reviewed.
Owner-chat feedback capture requires consent and retention controls. Capturing a review sample is separate from approving it for training. Your team reviews corrections before they shape another candidate.
02 / Know how data is processed
Use content you have permission to process. Approve provider recipients, retention, and applicable contractual terms before sharing confidential or regulated information. Project access controls do not mean local-only processing.
| Workflow | Processing |
|---|---|
| Training & model answers | Configured platform training and inference workers process project data and prompts. Model artifacts are stored in cloud object storage. |
| Living Memory retrieval | Document chunks and search queries are sent to OpenAI for embeddings. Retrieved passages provide context to the answering model. |
| Guided AI interviews | Recent conversation content is sent to configured Anthropic or OpenAI services for follow-up generation. OpenAI can be used after an Anthropic failure, so both recipients need approval. |
| Public URL import | Fetching a URL contacts its website and exposes request metadata. Do not include confidential information in URLs or query strings. |
| Model evaluation | Candidate and judge models run on evaluation workers. Artifact staging can contact model repositories; evaluation artifacts should be treated as sensitive project data. |
03 / Rights and responsibility
Your source rights remain yours. Upload only material you have permission to process and train on. An adapter specializes a base model; it does not transfer ownership of that base model or remove its license obligations. Review the selected model’s license and your agreement for use and export rights.
Corrections have approve or reject decisions and revision tracking. Dataset preparation separates training examples from held-out evaluation examples. Preparation, training, and activation are separate steps.
Erasure workflows track status and source lineage. Removing a source can invalidate derived adapters and require clean retraining; deleting a document does not erase its influence from existing model weights. Retained records and backups have their own retention obligations.
Confirm permitted data classes, processing recipients, storage location, access, retention, erasure, and model licenses. Obtain required permissions and contracts for regulated or confidential information. AI outputs still require review by the people accountable for the work.
Tell us about the work and requirements that matter to your team.