GLiNER (Entity Extraction)

Zero-shot named entity recognition service. Extract entities from text without training, using any custom labels.

Port: 8093
Model: urchade/gliner_multi-v2.1

What it Does

GLiNER extracts named entities from text with zero-shot capability:

Zero-shot NER - Define custom entity types at runtime
Multilingual - Works across multiple languages
Flexible labels - No predefined entity types required

Use cases:

Extract people, organizations, locations from documents
Find custom entities (products, dates, amounts)
Build knowledge graphs from unstructured text

API Endpoints

Extract Entities

Endpoint: POST /predict

curl -X POST "http://gliner-server:8093/predict" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Apple Inc. was founded by Steve Jobs in Cupertino, California.",
    "labels": ["person", "organization", "location"],
    "threshold": 0.5
  }'

Request:

Parameter	Type	Description
`text`	string	Text to analyze
`labels`	array	Entity types to extract
`threshold`	number	Confidence threshold (0-1, default: 0.5)

Response:

{
  "entities": [
    {"text": "Apple Inc.", "label": "organization", "score": 0.95, "start": 0, "end": 10},
    {"text": "Steve Jobs", "label": "person", "score": 0.92, "start": 27, "end": 37},
    {"text": "Cupertino, California", "label": "location", "score": 0.88, "start": 41, "end": 62}
  ]
}

Preload Model

Force model loading (useful during startup):

curl -X POST "http://gliner-server:8093/preload"

Health Check

curl http://gliner-server:8093/health

Usage Example

import requests
 
text = """
The merger between Microsoft and Activision Blizzard was approved 
by the FTC on October 13, 2023. The deal was worth $68.7 billion.
"""
 
response = requests.post(
    "http://gliner-server:8093/predict",
    json={
        "text": text,
        "labels": ["company", "organization", "date", "money"],
        "threshold": 0.4
    }
)
 
for entity in response.json()["entities"]:
    print(f"{entity['label']}: {entity['text']} ({entity['score']:.2f})")

Output:

company: Microsoft (0.91)
company: Activision Blizzard (0.88)
organization: FTC (0.85)
date: October 13, 2023 (0.92)
money: $68.7 billion (0.89)

Configuration

Variable	Default	Description
`GLINER_MODEL`	`urchade/gliner_multi-v2.1`	Model name
`GLINER_THRESHOLD`	`0.5`	Default confidence threshold
`GLINER_LABELS`	-	Default labels (comma-separated)
`GLINER_DEVICE`	auto	`cuda:0`, `mps`, `cpu`
`GLINER_MAX_CHARS`	-	Truncate long inputs

Tips

Lower threshold (0.3-0.4) for recall, higher (0.6-0.8) for precision
Specific labels work better than generic ones
Batch processing - Send multiple texts for efficiency

FastEmbed (Sparse Vectors)Jina Embeddings