# Entity Extraction

# Entity Extraction

Entity extraction identifies meaningful spans of text (entities) and turns them into structured outputs. Examples:

• products and device names
• people and organizations
• identifiers (ticket IDs, asset tags)
• locations and topics (depending on your domain)

In Curiosity Workspace, extraction is typically configured via NLP pipelines and models, then optionally linked into the graph.

# Extraction vs linking

• Extraction finds entities in text.
• Linking connects extracted entities to graph nodes (or creates nodes) so you can:
  • navigate from text to entities
  • filter by entities
  • use entities as grounding context for AI workflows

# Common model types (conceptual)

• Dictionary/spotter models
  • match known terms (product catalog, customer names)
• Pattern models
  • capture structured forms (IDs, serial formats, codes)
• ML models
  • detect entities that cannot be enumerated (optional, domain-dependent)

# Recommended workflow

1. Start with extraction on a single high-value field (e.g., Summary).
2. Run experiments to evaluate:
   • precision (how many captures are correct?)
   • recall (how much is missed?)
3. Iterate on model coverage and exclusions.
4. Enable linking into the graph only when extraction is reliable enough.

# Common pitfalls

• High false positives: pattern models can over-capture; add constraints and test broadly.
• Ambiguous names: dictionary models need aliases and disambiguation strategy.
• No evaluation loop: extraction needs iteration with real examples.

# Next steps

• Add semantic retrieval to complement extraction: Embeddings
• Implement domain-specific extraction rules: Custom NLP Rules