Extract Entities

Extract Entities

Identify and extract structured information from unstructured text: people, places, organizations, topics, dates, and relationships between entities.

Purpose

Transform free-form text into structured entity data for:

• Building knowledge graphs
• Indexing and retrieval
• Pattern detection across documents
• Linking related content

Input Format

Accepts:

• Plain text (paragraphs, documents)
• Structured data with text fields
• Lists of text snippets

Output Format

Returns JSON structure:

{
  "people": ["Name1", "Name2"],
  "organizations": ["Org1", "Org2"],
  "locations": ["Place1", "Place2"],
  "topics": ["Topic1", "Topic2"],
  "dates": ["2025-01-15", "last week"],
  "key_concepts": ["Concept1", "Concept2"],
  "relationships": [
    {"subject": "Name1", "predicate": "works_at", "object": "Org1"},
    {"subject": "Topic1", "predicate": "relates_to", "object": "Topic2"}
  ]
}

Extraction Guidelines

Entity Categories

People: Full names, roles, pronouns with clear referents

• Include professional titles if mentioned
• Resolve pronouns when unambiguous

Organizations: Companies, institutions, projects, teams

• Include both formal and informal names
• Note parent/subsidiary relationships

Locations: Cities, countries, venues, virtual spaces

• Be specific when possible (not just "the office")

Topics: Domain areas, technologies, methodologies

• Extract at appropriate granularity (not too broad/narrow)
• Include synonyms if multiple terms used

Dates/Time: Absolute and relative temporal references

• Normalize when possible (ISO format for absolute dates)
• Preserve relative references ("next week", "recently")

Key Concepts: Abstract ideas, themes, goals

• Focus on novel or emphasized concepts
• Distinguish from general background

Relationships

Extract explicit and strongly implied relationships:

• Employment/affiliation
• Collaboration/partnership
• Causation/dependency
• Temporal ordering
• Hierarchical structure

Format: {subject, predicate, object} triples

Quality Standards

• Precision over recall: Only extract clear, confident entities
• Disambiguation: Use context to resolve ambiguous references
• Normalization: Consistent entity naming across text
• No hallucination: Never infer entities not present in source

Special Handling

Pronouns: Resolve only when antecedent is clear and recent Abbreviations: Expand on first use, preserve thereafter Implicit entities: Extract only if strongly implied by context Conflicting info: Note conflicts in relationships field

Parameters

Optional args dict can specify:

• entity_types: List of types to extract (default: all)
• include_confidence: Boolean, add confidence scores (default: false)
• max_entities_per_type: Limit results (default: unlimited)

Example

Input:

Sarah joined Anthropic last quarter to work on constitutional AI. 
She previously collaborated with researchers at DeepMind on alignment.

Output:

{
  "people": ["Sarah"],
  "organizations": ["Anthropic", "DeepMind"],
  "topics": ["constitutional AI", "alignment"],
  "dates": ["last quarter"],
  "key_concepts": ["alignment research"],
  "relationships": [
    {"subject": "Sarah", "predicate": "works_at", "object": "Anthropic"},
    {"subject": "Sarah", "predicate": "previously_at", "object": "DeepMind"},
    {"subject": "Sarah", "predicate": "works_on", "object": "constitutional AI"}
  ]
}