> ## Documentation Index
> Fetch the complete documentation index at: https://avala.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Autotag

> Reference for Avala's ML-assisted autotag feature for accelerating annotation workflows

Autotag uses machine learning models to automatically generate suggested tags and labels for your data. These auto-generated tags help accelerate annotation by pre-labeling items that can then be reviewed and corrected by annotators.

## Concept

When autotag is enabled for a project, ML models analyze your data and produce confidence-scored predictions. These predictions appear as suggested annotations that annotators can accept, modify, or reject. Autotag works with both image and object-level predictions.

## Image Prefix Queries

Filter items by image-level autotag predictions using the `image_prefix:` syntax:

```
image_prefix:weather_clear
image_prefix:scene_highway
image_prefix:time_of_day_night
```

Image prefix queries match items where the autotag model has predicted a specific image-level classification.

## Object Prefix Queries

Filter items by object-level autotag predictions using the `object_prefix:` syntax:

```
object_prefix:car
object_prefix:pedestrian
object_prefix:traffic_sign
```

Object prefix queries match items that contain at least one predicted object of the specified type.

## Score Ranges

Autotag predictions include a confidence score ranging from **-1 to 1**:

| Score Range     | Meaning                                                       |
| --------------- | ------------------------------------------------------------- |
| `0.8` to `1.0`  | High confidence — model is very certain                       |
| `0.5` to `0.8`  | Medium confidence — likely correct but should be verified     |
| `0.0` to `0.5`  | Low confidence — uncertain prediction                         |
| `-1.0` to `0.0` | Negative confidence — model predicts the tag does *not* apply |

Filter by confidence score to focus on predictions that need review:

```
image_prefix:weather_clear AND score >= 0.8
object_prefix:car AND score < 0.5
```

<Tip>
  Items with low confidence scores are good candidates for manual review, as they represent cases where the model is uncertain.
</Tip>

## Training Set Queries

Filter items by the training set used to generate autotag predictions:

```
training_set = "model_v2"
training_set = "detector_2025_01"
```

This is useful when multiple autotag models have been run on the same dataset and you want to compare or filter by a specific model version.

## Usage in the Annotation Workflow

### Enabling Autotag

1. Navigate to your project in Avala.
2. Go to **Settings → Autotag**.
3. Select the ML model to use for predictions.
4. Choose the data to run autotag on (full dataset or specific slices).
5. Click **Run Autotag** to start the prediction job.

### Reviewing Autotag Predictions

1. Open the annotation editor for an item with autotag predictions.
2. Suggested annotations appear with a distinct visual indicator.
3. For each suggestion:
   * **Accept** — Confirm the prediction as correct.
   * **Modify** — Adjust the annotation (resize, relabel, etc.).
   * **Reject** — Remove the incorrect prediction.
4. Save your review to finalize the annotations.

### Filtering by Autotag Status

Use the query language to find items based on their autotag review status:

```
autotag_status = "pending"
autotag_status = "reviewed"
autotag_status = "accepted"
```

## Best Practices

* **Start with high-confidence predictions** — Review items with scores above 0.8 first for quick wins.
* **Use low-confidence items for model improvement** — These edge cases are valuable for retraining.
* **Run autotag on new data incrementally** — Process new uploads as they arrive rather than waiting for large batches.
* **Compare model versions** — Use training set queries to evaluate whether a newer model performs better.