Struct LLMEntityExtractor 

pub struct LLMEntityExtractor { /* private fields */ }

LLM-based entity extractor that uses actual language model calls

Implementations

impl LLMEntityExtractor

pub fn new(ollama_client: OllamaClient, entity_types: Vec<String>) -> Self

Create a new LLM-based entity extractor

Arguments

• ollama_client - Ollama client for LLM inference
• entity_types - List of entity types to extract (e.g., [“PERSON”, “LOCATION”, “ORGANIZATION”])

pub fn with_temperature(self, temperature: f32) -> Self

Set temperature for LLM generation (default: 0.1)

pub fn with_max_tokens(self, max_tokens: usize) -> Self

Set maximum tokens for LLM generation (default: 1500)

pub fn with_keep_alive(self, keep_alive: Option<String>) -> Self

Set keep_alive for KV cache persistence (e.g. “1h”, “30m”).

pub fn estimate_tokens(text: &str) -> u32

Estimate token count from text length (≈ 1 token per 4 chars).

pub fn calculate_entity_num_ctx( built_prompt: &str, max_output_tokens: u32, ) -> u32

Calculate the required num_ctx for one entity-extraction LLM call.

Formula (mirrors ContextualEnricher::calculate_num_ctx):

num_ctx = tokens(built_prompt)   // instructions + entity types + chunk text
        + max_output_tokens       // JSON with entities + relationships
        + 20% safety margin       // avoids silent truncation

Result is rounded up to the nearest 1024 and clamped to [4096, 131072].

pub async fn extract_from_chunk( &self, chunk: &TextChunk, ) -> Result<(Vec<Entity>, Vec<Relationship>)>

Extract entities and relationships from a text chunk using LLM

This is the REAL extraction that makes actual LLM API calls. Expected time: 15-30 seconds per chunk depending on chunk size and model.

pub async fn extract_additional( &self, chunk: &TextChunk, previous_entities: &[EntityData], previous_relationships: &[RelationshipData], ) -> Result<(Vec<Entity>, Vec<Relationship>)>

Extract additional entities in a gleaning round (continuation)

This is used after the initial extraction to catch missed entities.

pub async fn check_completion( &self, chunk: &TextChunk, entities: &[EntityData], relationships: &[RelationshipData], ) -> Result<bool>

Check if extraction is complete using LLM judgment

Uses the LLM to determine if all significant entities have been extracted.