Model Repository
The Model Repository is the catalog-backed inventory for AI and machine learning assets in Qarion. It extends data products with model-specific metadata, versions, and artifacts so model owners can keep governance, operational state, and release files in the same place as lineage, quality, risk, and stewardship context.
Use the Model Repository for model records and model artifacts. It is separate from engagement Git repositories and from general artifact repositories used for Python packages, OCI images, Docker images, or Helm charts.
When To Use It
Use the Model Repository when you need to:
- Register a machine learning model, AI system, agent, AI app, or prompt collection as a governed catalog product.
- Track model lifecycle, deployment status, risk classification, intended use, limitations, framework, architecture, and serving information.
- Compare model versions and mark the current release.
- Attach artifacts such as binaries, weights, checkpoints, adapters, tokenizers, prompt bundles, evaluation reports, and model cards.
- Keep checksums and upload status visible for large model files.
Use Artifact Repositories when the asset is a package registry or container repository rather than a model record.
Scope And Discovery
The repository can be viewed at two scopes:
| Scope | Use it for |
|---|---|
| Space | Create and manage model records inside one governed space. |
| Organization | Browse model records across spaces that you can access. |
Both scopes support filtering by search text, model family, product type, deployment status, task type, risk classification, and archived state.
Organization scope is read-oriented discovery. Create model records from the space where the model should live so ownership, permissions, quality signals, and governance metadata stay attached to the correct space.
Supported Product Types
Model Repository entries are data products with AI and ML product types.
| Product type | Use it for |
|---|---|
| ML Model | Trained predictive, ranking, embedding, forecasting, or analytical models. |
| AI System | AI systems registered for governance, compliance, and lifecycle tracking. |
| LLM Agent | Conversational or tool-using large language model agents. |
| AI App | Applications or services whose primary behavior is AI-powered. |
| Agent | Autonomous or semi-autonomous agentic workflows. |
| Prompt Collection | Governed prompt libraries and prompt bundles. |
For the full catalog behavior of these product types, see AI and ML product types.
Model Families And Task Types
Model family drives repository filtering and helps users understand the asset class at a glance.
| Family | Typical use |
|---|---|
traditional_ml | Classical ML models such as tree-based models, regressors, and classifiers. |
genai | Generative AI models and systems. |
foundation_model | Base or foundation models used directly or adapted by other systems. |
embedding_model | Embedding models used for search, retrieval, similarity, or ranking. |
agent | Agentic systems with tool use, planning, or multi-step behavior. |
ai_app | AI-powered applications and services. |
prompt_collection | Prompt libraries, prompt packs, and prompt bundles. |
other | Model assets that do not fit the standard families. |
Task type captures what the model does. Common values include llm,
classification, regression, clustering, forecasting, embedding,
ranking, anomaly_detection, recommendation, agent, and other.
Creating A Model Entry
Create the model from the repository view in the target space. At minimum, provide a name. Qarion can also store:
- Slug, description, and long-form description.
- Product type and provider.
- Environment, lifecycle stage, deployment status, and serving URL.
- Risk classification for governance and AI Act-oriented review.
- Model family, task type, architecture, framework, intended use, and limitations.
- Input and output schema as JSON metadata.
- Vendor, vendor product, vendor model ID, deployment environment, and business use case for externally supplied systems.
- GPAI indicator and AI system type where applicable.
Model Repository records remain catalog products. Ownership, stewardship, comments, attachments, lineage, quality checks, access workflows, and governance metadata continue to live in the standard product experience.
Versions
Use the versions tab on a model record to track release history. Versions can represent releases, major/minor/patch changes, milestones, alpha builds, or beta builds. A version can include release notes, a release date, and source control metadata such as a Git tag or commit SHA when that context is available.
The repository summary highlights version count and current version information so users can quickly see whether the model has a documented release history.
Artifacts
Artifacts are files or external artifact references attached to a model product. They are listed on the artifacts tab and can optionally be tied to a product version or training run.
| Artifact kind | Examples |
|---|---|
model_binary | Serialized model package or export. |
weights | Weight files such as safetensors or framework checkpoints. |
checkpoint | Training or fine-tuning checkpoint. |
adapter | Adapter or LoRA-style delta artifact. |
tokenizer | Tokenizer files and vocabulary assets. |
prompt_bundle | Prompt templates or prompt packs. |
eval_report | Evaluation output, benchmark report, or validation report. |
model_card | Model card or release documentation. |
other | Supporting artifact that does not fit the standard kinds. |
Each artifact stores file name, content type, size, SHA-256 checksum, artifact kind, optional version label, upload status, uploader, timestamps, storage location metadata, and whether it is the current artifact for the model.
Chunked Uploads And Checksums
Managed uploads are designed for large model files:
- Start an upload session with file name, content type, total size, artifact kind, and optional expected SHA-256 checksum.
- Upload the file in numbered chunks using the chunk size returned by the session.
- Complete the session. Qarion assembles the chunks, computes SHA-256, stores the artifact, and records the final checksum.
If you provide an expected checksum at upload initiation or completion, the computed checksum must match before the artifact is accepted. The checksum is shown on the artifact so downstream users can verify the downloaded file.
Upload sessions expire if they are not completed in time. The maximum upload size, chunk size, session TTL, and temporary staging directory are controlled by platform configuration.
Current Artifact Behavior
Artifacts can be marked current. Use this to identify the artifact that should be treated as the active release for the model, while retaining older artifacts for comparison, audit, or rollback context.
Current artifact fields appear in the repository summary, including artifact ID, file name, kind, content type, detected model format, SHA-256 checksum, and file size. Deleting an artifact removes it from the active artifact list; keep model cards and evaluation reports as separate artifacts when you need those records to remain independently discoverable.
Governance
Because models are catalog products, repository records participate in the same governance model as other data products:
- Owners, stewards, custodians, tags, and product documentation remain attached to the product.
- Risk classification and AI-specific metadata support AI governance reviews.
- Lineage can connect training data, models, and downstream consumers.
- Quality checks and monitoring can be used to track model or data health where those checks are configured.
- Change requests and approval workflows can be used for governed model updates.
See Product Details and AI Governance for the adjacent governance workflows.
Troubleshooting
| Symptom | What to check |
|---|---|
| A model is missing from organization scope. | Confirm you have access to the model's space and that filters are not excluding the product type, family, task type, deployment status, risk class, or archived state. |
| You cannot create a model. | Confirm you are in space scope, have space access, and the space has not reached its catalog product limit. |
| You cannot edit metadata or upload artifacts. | Confirm you have edit access to the model product. Artifact listing and downloads require view access; upload and delete require edit access. |
| Upload completion fails with a checksum error. | Recompute the local SHA-256 digest and retry completion with the expected checksum that matches the uploaded bytes. |
| Upload completion fails after a long delay. | Start a new upload session; incomplete sessions expire after the configured TTL. |
| A download returns no file. | Confirm the artifact was uploaded successfully and has not been deleted. External storage can also return a redirect instead of a local file response. |