---
title: How to extract entities and structured data with SIE
description: "SIE's extract primitive pulls structured information from unstructured text and images. It handles named entity recognition, relation extraction, text classification, and vision tasks like OCR and captioning."
canonical_url: https://superlinked.com/docs/extract
last_updated: 2026-05-20
---

**SIE's `extract` primitive pulls structured information from unstructured content.** It handles named entity recognition (NER), relation extraction, text classification, and vision tasks including captioning and OCR. Models run on your own infrastructure with zero per-call API costs.

#### Python

```python
from sie_sdk import SIEClient
from sie_sdk.types import Item

client = SIEClient("http://localhost:8080")

text = Item(text="Apple CEO Tim Cook announced the iPhone 16 in Cupertino.")
result = client.extract(
    "urchade/gliner_multi-v2.1",
    text,
    labels=["person", "organization", "product", "location"]
)
for entity in result["entities"]:
    print(f"{entity['label']}: {entity['text']} (score: {entity['score']:.2f})")
# organization: Apple (score: 0.95)
# person: Tim Cook (score: 0.93)
# product: iPhone 16 (score: 0.89)
# location: Cupertino (score: 0.87)
```

#### TypeScript

```typescript
import { SIEClient } from "@superlinked/sie-sdk";

const client = new SIEClient("http://localhost:8080");

const result = await client.extract(
  "urchade/gliner_multi-v2.1",
  { text: "Apple CEO Tim Cook announced the iPhone 16 in Cupertino." },
  { labels: ["person", "organization", "product", "location"] },
);
for (const entity of result.entities) {
  console.log(`${entity.label}: ${entity.text} (score: ${entity.score.toFixed(2)})`);
}
// organization: Apple (score: 0.95)
// person: Tim Cook (score: 0.93)
// product: iPhone 16 (score: 0.89)
// location: Cupertino (score: 0.87)

await client.close();
```

For model recommendations, see the full [model catalog](https://superlinked.com/models#task=extract).

---

## Input Types

`Item` accepts three input modes depending on the model:

- **`text`** — plain string. Used by GLiNER, GLiREL, GLiClass, and the rest of the text-only extractors.
- **`images`** — list of image bytes (or `{data, format}` dicts in Python). Used by Florence-2, Donut, GroundingDINO, OWL-v2, and image-input OCR models like `zai-org/GLM-OCR`, `lightonai/LightOnOCR-2-1B`, and `PaddlePaddle/PaddleOCR-VL-1.5`. See [Vision Tasks](https://superlinked.com/docs/extract/vision/) and [OCR](https://superlinked.com/docs/extract/ocr/).
- **`document`** — raw file bytes (PDF, DOCX, HTML, MD, TXT, RTF, ODT, PPTX, XLSX, CSV). Used by the multi-page `docling` parser. The Python SDK auto-detects the format from a path suffix; bytes-based callers pass `format` explicitly. See [OCR → Docling](https://superlinked.com/docs/extract/ocr/#docling-multi-page-documents).

---

## Named Entity Recognition

GLiNER models support zero-shot NER: define any entity types you need at query time, with no predefined schema.

### Custom Entity Types

#### Python

```python
result = client.extract(
    "urchade/gliner_multi-v2.1",
    Item(text="The merger between Acme Corp and Beta Inc requires FTC approval."),
    labels=["company", "regulatory_body", "legal_action"]
)
for entity in result["entities"]:
    print(f"{entity['label']}: {entity['text']}")
# company: Acme Corp
# company: Beta Inc
# regulatory_body: FTC
```

#### TypeScript

```typescript
const result = await client.extract(
  "urchade/gliner_multi-v2.1",
  { text: "The merger between Acme Corp and Beta Inc requires FTC approval." },
  { labels: ["company", "regulatory_body", "legal_action"] },
);
for (const entity of result.entities) {
  console.log(`${entity.label}: ${entity.text}`);
}
// company: Acme Corp
// company: Beta Inc
// regulatory_body: FTC
```

### Entity Positions

Each entity includes character positions for highlighting or downstream processing:

#### Python

```python
result = client.extract(
    "urchade/gliner_multi-v2.1",
    Item(text="Tim Cook works at Apple."),
    labels=["person", "organization"]
)
for entity in result["entities"]:
    print(f"{entity['label']}: '{entity['text']}' [{entity['start']}:{entity['end']}]")
# person: 'Tim Cook' [0:8]
# organization: 'Apple' [18:23]
```

#### TypeScript

```typescript
const result = await client.extract(
  "urchade/gliner_multi-v2.1",
  { text: "Tim Cook works at Apple." },
  { labels: ["person", "organization"] },
);
for (const entity of result.entities) {
  console.log(`${entity.label}: '${entity.text}' [${entity.start}:${entity.end}]`);
}
// person: 'Tim Cook' [0:8]
// organization: 'Apple' [18:23]
```

### Batch Extraction

#### Python

```python
documents = [
    Item(id="doc-1", text="Microsoft acquired Activision for $69 billion."),
    Item(id="doc-2", text="Sundar Pichai leads Google's AI initiatives."),
]
results = client.extract(
    "urchade/gliner_multi-v2.1",
    documents,
    labels=["person", "organization", "money"]
)
```

#### TypeScript

```typescript
const documents = [
  { id: "doc-1", text: "Microsoft acquired Activision for $69 billion." },
  { id: "doc-2", text: "Sundar Pichai leads Google's AI initiatives." },
];
const results = await client.extract(
  "urchade/gliner_multi-v2.1",
  documents,
  { labels: ["person", "organization", "money"] },
);
```

---

## Response Format

The `ExtractResult` contains different fields depending on the extraction type used:

| Field | Type | When present |
| --- | --- | --- |
| `id` | `str or None` | Always (if provided in input) |
| `entities` | `list[Entity]` | NER models (GLiNER) |
| `relations` | `list[Relation]` | Relation extraction (GLiREL) |
| `classifications` | `list[Classification]` | Classification models (GLiClass) |
| `objects` | `list[DetectedObject]` | Object detection (GroundingDINO, OWLv2) |
| `data` | `dict` | Document/composite extractors (Docling, Donut, document-mode Florence-2) |

### Entity Fields

| Field | Type | Description |
| --- | --- | --- |
| `text` | `str` | Extracted text span |
| `label` | `str` | Entity type label |
| `score` | `float` | Confidence score from 0 to 1 |
| `start` | `int` | Start character position |
| `end` | `int` | End character position |

---

## HTTP API

The server defaults to msgpack. For JSON responses:

```bash
curl -X POST http://localhost:8080/v1/extract/urchade/gliner_multi-v2.1 \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{
    "items": [{"text": "Tim Cook is the CEO of Apple."}],
    "params": {"labels": ["person", "organization"]}
  }'
```

See the [HTTP API Reference](https://superlinked.com/docs/reference/api/).

---

## Framework Integrations

Extraction is available through all major framework integrations, not just the native SDK:

| Framework | Component | Returns |
| --- | --- | --- |
| [LangChain](https://superlinked.com/docs/integrations/langchain/) | `SIEExtractor` | Dict with `entities`, `relations`, `classifications`, `objects` |
| [LlamaIndex](https://superlinked.com/docs/integrations/llamaindex/) | `create_sie_extractor_tool` | Dict with `entities`, `relations`, `classifications`, `objects` |
| [Haystack](https://superlinked.com/docs/integrations/haystack/) | `SIEExtractor` | Typed outputs: `Entity`, `Relation`, `Classification`, `DetectedObject` |
| [DSPy](https://superlinked.com/docs/integrations/dspy/) | `SIEExtractor` | `dspy.Prediction` with extraction fields |
| [CrewAI](https://superlinked.com/docs/integrations/crewai/) | `SIEExtractorTool` | Formatted string with all extraction types |

---

## Frequently Asked Questions

**What is zero-shot NER?**
Zero-shot NER means you can define your entity types at query time without fine-tuning a model. GLiNER models like `urchade/gliner_multi-v2.1` accept arbitrary label strings and extract matching spans from text. There is no fixed list of entity types.

**Does SIE support relation extraction?**
Yes. GLiREL models extract relationships between entities (for example, "CEO of", "acquired by"). See [Relations and Classification](https://superlinked.com/docs/extract/relations/).

**Can SIE extract data from PDFs and images?**
Yes. SIE supports four dedicated OCR models: `zai-org/GLM-OCR`, `lightonai/LightOnOCR-2-1B`, `PaddlePaddle/PaddleOCR-VL-1.5`, and `docling` (multi-page PDF/DOCX/HTML). They convert documents to Markdown while preserving tables and layout. Donut and Florence-2 are also available for image captioning and visual QA. See [OCR](https://superlinked.com/docs/extract/ocr/) and [Vision Tasks](https://superlinked.com/docs/extract/vision/).

**Which model should I use for entity extraction?**
`urchade/gliner_multi-v2.1` is a strong default for multilingual NER. It handles zero-shot extraction across 100+ languages. Browse all extraction models in the [model catalog](https://superlinked.com/models#task=extract).