> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.ovalbee.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.ovalbee.com/_mcp/server.

# Node library

Each page in the Node Library describes one [node](/concepts/nodes): what it does and when to reach for it, what has to exist before it runs, what happens step by step, answers to common questions, and its exact inputs and outputs. Open a node's page from the app with the **Documentation** link in the node's sidebar.

## Categories

The sidebar groups nodes the same way the node picker in the editor does.

| Category         | What its nodes do                                                                                                                          |
| ---------------- | ------------------------------------------------------------------------------------------------------------------------------------------ |
| **Import**       | Bring images and videos into a dataset: upload files or pull a public dataset.                                                             |
| **Datasets**     | Point a flow at a dataset, copy, move and merge data, create ontologies, and profile a dataset's quality.                                  |
| **Assets**       | Remove assets and transform annotations: keep some classes, one shape per class, rename classes.                                           |
| **Filter**       | Keep or drop images, labels and predictions by rules.                                                                                      |
| **Sample**       | Take random samples, split into training and validation, collect images worth a second look.                                               |
| **Label**        | Labeling and review stages, and automated quality checks between them.                                                                     |
| **Train**        | Fine-tune detectors on your data.                                                                                                          |
| **Predict**      | Run a pretrained or trained detector over a dataset.                                                                                       |
| **Evaluate**     | Score predictions against ground truth, compare models, recommend training settings.                                                       |
| **Models**       | Export and upload models, and embedding-based tools for duplicates, outliers, label errors and dataset maps.                               |
| **Flow control** | Start a chain of nodes, wait for branches to finish, stop a chain, or let a language model decide from earlier reports whether it goes on. |
| **Supervisely**  | Import from Supervisely and run Supervisely labeling queues and jobs from a flow.                                                          |

## Reading a node page

* **Badges** at the top name the kinds of machine a node runs on — `Runs on: gpu` needs a GPU machine class, `Runs on: gpu, cpu` takes either — and mark nodes that run in their own **Docker** image, whose first run can take longer while the image is pulled.
* **Prerequisites** list what must already exist, such as a labeled dataset or a secret with credentials.
* **Inputs** and **Outputs** are generated from the node's definition, so they always match what you see in the sidebar. The **JSON config** at the bottom is the same interface in machine-readable form.

Looking for the ideas behind the nodes — how they pass data, when they run? Start with [References](/concepts/references) and [Runs and automation](/concepts/runs).