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

# Advanced Usage

> Restrict which codes the model may predict, and how filter validation works

### Filter the prediction set

Use the optional `filter` field to restrict which codes the model may predict. `include` and `exclude` each take a list of condition objects:

* **`include`**: codes matching all of these conditions are eligible. Empty or omitted means every code in the system is eligible.
* **`exclude`**: codes matching all of these conditions are removed from the eligible set. Empty or omitted means nothing is removed.

The final prediction set is the include set minus the exclude set.

Each condition has the following fields:

| Parameter  | Description                                                                                   |
| ---------- | --------------------------------------------------------------------------------------------- |
| `property` | The attribute to filter on, for example `code` (matches the `code` field on predicted codes). |
| `op`       | Comparison operator, one of `=`, `is-a`, `descendent-of`, `exists`, `in`. Defaults to `=`.    |
| `value`    | Comparison value.                                                                             |

`op` supports:

| Operator        | Matches                                                                                                         |
| --------------- | --------------------------------------------------------------------------------------------------------------- |
| `=`             | The property equals `value`.                                                                                    |
| `is-a`          | The code itself or any of its descendants.                                                                      |
| `descendent-of` | A strict descendant of the code (excludes the code itself).                                                     |
| `exists`        | Whether the property is set. `value: true` matches when it has a value; `value: false` matches when it doesn't. |
| `in`            | The property's value is one of the values in `value`.                                                           |

For example, to predict type 2 diabetes mellitus (`E11`) and its descendants, but never the specific code for diabetes with hyperosmolarity (`E11.00`):

```json theme={null}
"filter": {
  "include": [
    { "property": "code", "op": "is-a", "value": ["E11"] }
  ],
  "exclude": [
    { "property": "code", "op": "=", "value": "E11.00" }
  ]
}
```

<Tip>
  Conditions in `include`/`exclude` are AND-combined. To match alternatives, list them in one condition's `value` instead of adding separate conditions, for example `op: "in"` for exact values or `op: "is-a"` with multiple ancestor codes where supported.
</Tip>

Some coding systems support filtering on additional attributes beyond `code`. For example, SNOMED CT concepts can be filtered by `semantic_tag`, the clinical category a concept belongs to:

```json theme={null}
"filter": {
  "include": [
    { "property": "semantic_tag", "op": "in", "value": ["disorder", "finding"] }
  ]
}
```

See [SNOMED CT](/coding/snomed-ct#filter-by-attribute) for the full list of supported `semantic_tag` values.

#### Filter validation

Filter conditions are validated before prediction runs. A request is rejected with a 400 error if:

* `property` isn't a supported attribute for the coding system (`code`, or a system-specific attribute like `semantic_tag`).
* `value` doesn't match the type `op` expects — a string or list of strings for `is-a`/`descendent-of`, a boolean for `exists`, a list for `in`.
* `is-a` or `descendent-of` is used on a property other than `code`.
* A code referenced in a condition (directly, or as the ancestor in an `is-a`/`descendent-of` condition) doesn't exist in the coding system.

When requesting multiple coding systems, the same `filter` applies to all of them. If a condition doesn't apply to one of the requested systems — for example, a `semantic_tag` condition when also requesting CPT — that system fails validation and the entire request is rejected, not just that system.

No prediction is attempted when validation fails. See [Errors](/api-reference/errors) for the general error response shape.

<Tip>
  Omit the `filter` field for open-ended prediction: the API will predict across the full coding system.
</Tip>

<Note>Please [contact us](mailto:help@corti.ai) if you need help configuring your requests or have questions about advanced usage.</Note>