Extract-it
The extract_inference tool allows LLMs and autonomous agents to extract open-vocabulary named entities, spans, confidence scores, and relationships from unstructured text without hallucinations.
Tool Overview
- Tool Name:
extract_inference - Underlying Endpoint:
POST /inference/extract/ - SDK Methods:
kg.extract.it(Python) /kg.extract.it(TypeScript)
LoRA Adapters vs. Open Models
adapter_id is for fine-tuned LoRA adapters attached to the base Extract-it service. If you trained or deployed a standalone Open Model (e.g., GLiClass or GLiNER checkpoint), use the open_model_inference tool instead.
Parameters
| Parameter | Type | Required | Description |
|---|---|---|---|
text | String | Yes | Target input text to extract entities or relations from. |
tasks | Object / Dictionary | Yes | Dictionary of extraction tasks following the Unified Inference API (e.g. {"entities": {"type": "span", "labels": ["company", "revenue"], "threshold": 0.4}}). |
adapter_id | String | No | Optional LoRA adapter ID for fine-tuned adapters (attaches KnowledgatorAPI-Model-ID header). |
timeout_ms | Integer | No | Request timeout in milliseconds (default: 60000 = 60s). |
Example Usage
Natural Language Prompt to Agent
"Extract company names and monetary amounts from: 'Alphabet Inc. reported Q4 revenues reaching $86.3 billion, surpassing Wall Street expectations.'"
Tool Call Arguments
{
"name": "extract_inference",
"arguments": {
"text": "Alphabet Inc. reported Q4 revenues reaching $86.3 billion, surpassing Wall Street expectations.",
"tasks": {
"entities": {
"type": "span",
"labels": ["company", "revenue", "reporting_date"],
"threshold": 0.4
}
}
}
}
Tool Response
{
"outputs": {
"entities": {
"type": "span",
"value": [
{
"span": "Alphabet Inc.",
"entity": "company",
"start": 0,
"end": 13,
"score": 0.985
},
{
"span": "$86.3 billion",
"entity": "revenue",
"start": 44,
"end": 57,
"score": 0.962
}
],
"threshold": 0.4
}
}
}