> For the complete documentation index, see [llms.txt](https://raic-labs.gitbook.io/raic-fed/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://raic-labs.gitbook.io/raic-fed/analyst-guide/03f-building-a-shallow-model.md).
# Building a shallow model
Shallow models are lightweight detectors built directly from taxonomy categories. They adapt from best-in-class models and are optimized for fast iteration and rapid subclassing, requiring few labeled examples and no AI specialist to build or refine.
Once a taxonomy category contains enough approved images, the Analyst can build a shallow model from it, train it, run it against imagery, and feed positive results back into the taxonomy. Each iteration improves precision without starting over.
***
## The Models section
The Models section is accessible from the **sparkle icon** in the left navigation menu. It displays all models in the shared workspace, with the following columns:
* **Model Name** — the name assigned when the model was created.
* **Latest Version** — the current version number.
* **Categories** — the taxonomy categories the model was built from.
* **Status** — Draft, Train, Failed or Ready.
* **Created** — when the model was first created.
* **Version Created** — when the current version was created.
* **Last Run** — when the model was last run against imagery.
Selecting a model in the list opens its detail panel on the right, showing source categories, version status, and run history.
***
## Step 1: Build a new model
1. Select **Build New** in the top right of the Models section.
2. In the **Build New Model** panel, enter a **Model Name**.
3. Under **Select Categories**, expand the taxonomy tree and check the categories to build the model from. The button at the bottom updates to show the number of images selected.
4. Select **Build New Model with \[n] Images** to create the model.
The model appears in the list with a status of **Draft** while processing. Once processing is complete the status updates to **Ready**.
> **Analyst** Negative samples are generated automatically. The Analyst does not need to provide them.
***
## Step 2: Train the first version
Once the model status is **Ready**, the Analyst trains the first version.
1. Select the model from the list to open its detail panel.
2. Select **Train** to build the first version.
3. The version status will show as **Draft** while training. Once training is complete it updates to **Ready**.
***
## Step 3: Run the model against imagery
Once a model version has a status of **Ready**, the Analyst can run it against a single image file.
1. Select the model from the list to open its detail panel.
2. Confirm the correct version is selected in the version dropdown.
3. Select **Run**.
4. Select the Detector used to ingest the imagery you would like to run the model against.
5. Find the image to run model on by scrolling the list, using paging at the bottom of the list, or filtering by the file name.
6. Select the image by checking box next the image name.
7. Choose Run Model.
***
## Step 4: Viewing model results
To view the results of a completed run, select the **eye icon** next to the run in the Run History table. Results display as bounding boxes overlaid on the source imagery in the map view. Each detection shows a confidence score.
The following controls are available in the results view:
* **Minimum Confidence** slider at the bottom of the screen filters results by confidence threshold. Adjust it to show only detections above a chosen confidence level.
* **Show Confidence** toggle displays or hides confidence scores on the bounding boxes.
* **Categories** filter at the bottom left allows the Analyst to filter results by taxonomy category.
* The total detection count is shown in the Search Scope bar at the top.
***
## Step 5: Adding results to the taxonomy
Detections from a model run can be added back into the taxonomy to further refine future model versions.
1. In the results view, select detections to add.
2. In the left panel, select the taxonomy category to assign them to.
3. Select **Save Detections to Category** to confirm.
Returning positive results to the taxonomy creates a feedback loop: each model run produces better-labeled examples, which improve the next version of the model.
***
## Creating a new model version
When the taxonomy changes, the Analyst can create a new model version to reflect those changes.
1. Select the model from the list to open its detail panel.
2. Select **New Version**.
3. In the **Create New Version** panel, review or adjust the selected categories.
4. Select **Create New Version with \[n] Images** to build.
The new version appears in the version dropdown once training is complete. Previous versions are retained and can be selected for comparison or re-running.
## Exporting results
After the model is run against an image, the Analyst can export results in a geojson format to use outside of the RAIC platform. Results include displayed detections, matching the confidence threshold and category filters. Export includes class, confidence, geographic coordinates, timestamp reference, source file name, crop url, and model metadata.
1. Select **Export Results** in the top right of the results view to export model detections.
2. You will be notified when the exported file is downloaded to your loca drive.
3. Export results are in geojson format.
{% embed url="" %}
***
## What's next
With shallow models producing reliable detections, the Analyst can continue iterating: refining taxonomy categories, creating new model versions, and running searches seeded from curated results. Each cycle builds a more precise, mission-specific picture of what is in the imagery.
**Compliance Note** This documentation does not constitute legal or regulatory guidance. Organizations operating in classified or regulated environments should consult their compliance and security teams before deploying or using RAIC Fed. Specific data handling, classification, and audit requirements will vary by environment and mission.