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

# AI Detection

> How Nvisy recognizes sensitive data in documents

Detection combines several recognition strategies and reconciles their results
into a single set of entities, each with a label, confidence, and location.

## Recognition Strategies

<CardGroup cols={2}>
  <Card title="Patterns" icon="regex">
    Deterministic matching for structured values
  </Card>

  <Card title="NER" icon="brain">
    Named-entity recognition for names, places, and organizations
  </Card>

  <Card title="LLM" icon="microchip">
    Model-driven classification for context-dependent data
  </Card>

  <Card title="Vision" icon="image">
    Recognition over rendered pages and embedded images
  </Card>
</CardGroup>

Recognition is deployment-wide: the built-in pattern set plus whichever
recognizers the engine has registered. Pipelines choose *what* to target and
*how* to treat it, not which recognizers exist.

## The Catalog

Two read-only endpoints describe what a deployment offers:

```
GET /catalog/labels/
GET /catalog/recognizers/
```

`labels` returns the built-in taxonomy — the categories of sensitive data
(PII, PHI, PCI, and so on) that policies can target. `recognizers` returns the
engine's registered lineup, grouped into `ner` and `llm`.

```typescript theme={null}
const labels = await client.catalog.listLabels();
const recognizers = await client.catalog.listRecognizers();

console.log(recognizers.ner.length, recognizers.llm.length);
```

<Tip>
  Query the catalog rather than hardcoding label names. It reflects what the
  deployment you are talking to actually supports, which differs between cloud
  and self-hosted installs.
</Tip>

## LLM Providers

Model-driven recognition runs against a connected provider — `openai`,
`anthropic`, or `ollama` for self-hosted inference. Configure one as a
workspace [connection](/features/integrations).

<Note>
  With `ollama`, inference stays inside your infrastructure. Combined with an
  on-premise deployment, no document content leaves your network.
</Note>

## Reconciliation

Multiple recognizers frequently find the same thing, or disagree. Rather than
taking the first match, the engine reconciles them — and records how. Each
entity's audit chain shows the steps applied:

| Event           | Meaning                              |
| --------------- | ------------------------------------ |
| `pattern`       | Matched a deterministic pattern      |
| `model`         | Recognized by a model                |
| `deduplication` | Merged with an overlapping detection |
| `conflict`      | Competing detections were reconciled |
| `contested`     | Recognizers disagreed                |
| `calibration`   | Confidence was calibrated            |
| `refinement`    | Boundaries or label were refined     |

Events carry their `parents`, so a final entity can be traced back through
every merge that produced it. See [Audit Trails](/features/audit-trails).

## Entities

Each recognized entity records:

| Field        | Description                                            |
| ------------ | ------------------------------------------------------ |
| `label`      | The taxonomy label it was classified as                |
| `confidence` | Calibrated confidence                                  |
| `location`   | Position, in coordinates native to the modality        |
| `language`   | Detected language, where applicable                    |
| `coref`      | Coreference link to other mentions of the same subject |

<Note>
  Coreference links matter for redaction consistency: when the same person is
  mentioned several ways, linked mentions can be treated as one subject rather
  than as unrelated findings.
</Note>

## Scope

A run's scope narrows recognition to the languages and jurisdictions that
apply, and controls OCR:

```json theme={null}
{
  "scope": {
    "languages": ["en"],
    "countries": ["US"],
    "ocrMode": { "kind": "force", "dpi": 300 }
  }
}
```

Pipelines carry a default scope; individual runs may override it.

## Review

Detection never modifies a document. A run stops at `analyzed` with its
findings recorded, and redaction is a separate, explicit step — so confidence
scores inform a human decision rather than replacing one.

## Next Steps

<CardGroup cols={2}>
  <Card title="Redaction Workflow" icon="workflow" href="/features/redaction-workflow">
    What happens after detection
  </Card>

  <Card title="Integrations" icon="plug" href="/features/integrations">
    Connect an LLM provider
  </Card>
</CardGroup>