Behaviour for entity extraction in GraphRAG.

Entity extractors identify named entities (people, organizations, locations, etc.) from text. Arcana provides a built-in NER implementation, but you can implement custom extractors for different approaches.

Built-in Implementations

• Arcana.Graph.EntityExtractor.NER - Local Bumblebee NER (default)

Configuration

Configure your entity extractor in config.exs:

# Default: Local NER with distilbert-NER
config :arcana, :graph,
  entity_extractor: :ner

# Custom module implementing this behaviour
config :arcana, :graph,
  entity_extractor: MyApp.LLMEntityExtractor

# Custom module with options
config :arcana, :graph,
  entity_extractor: {MyApp.LLMEntityExtractor, model: "gpt-4"}

# Inline function
config :arcana, :graph,
  entity_extractor: fn text, opts -> {:ok, extract_entities(text)} end

Implementing a Custom Extractor

Create a module that implements this behaviour:

defmodule MyApp.LLMEntityExtractor do
  @behaviour Arcana.Graph.EntityExtractor

  @impl true
  def extract(text, opts) do
    llm = opts[:llm] || raise "LLM required"
    # Use LLM to extract entities...
    {:ok, entities}
  end

  # Optional: implement for batch optimization
  @impl true
  def extract_batch(texts, opts) do
    # Batch LLM call...
    {:ok, results}
  end
end

Entity Format

Extractors must return entities as maps with at least:

• :name - The entity name (required)
• :type - Entity type as atom: :person, :organization, :location, :concept, :other

Optional fields:

• :span_start - Character offset where entity starts
• :span_end - Character offset where entity ends
• :score - Confidence score (0.0-1.0)
• :description - Brief description of the entity