LLM Extraction

LLM extraction turns source content into typed edges for graph enrichment. Memory entities are embedded for recall by default, but graph extraction for memory is opt-in.

What this page covers

Provider selection, model configuration, and practical flags for tuning graph quality, including memory-specific extraction policy.

Core workflow

1. For Docker Compose, set extraction in Admin Config, then save, validate, and apply it. For a non-Compose local process, set the environment before starting the server:

export EXTRACTION_ENABLED=true

2. Choose provider/model:

export EXTRACTION_PROVIDER=openai
export EXTRACTION_MODEL=gpt-4o-mini

3. Leave memory extraction in semantic-recall mode unless you explicitly want memory-derived graph edges:

export EXTRACTION_MEMORY_MODE=embed_only

Use extract_durable to extract only durable_memory, or extract_all to extract both durable memory and session context.

4. Run graph refresh if relationships were sparse:

./bin/pgm-admin reextract --type document --no-edges-only

5. Tune neighbor expansion when you need thematic links beyond explicit mentions:

export EXTRACTION_SEMANTIC_NEIGHBORS_ENABLED=true
export EXTRACTION_SEMANTIC_NEIGHBORS_MIN_SIMILARITY=0.75

Notes

• Supported providers: openai, anthropic, ollama, openai-compatible.
• Optional semantic neighbors create related_to edges from embedding similarity.
• Auto-creation is controlled via EXTRACTION_AUTO_CREATE_ENTITIES and type allowlists.
• EXTRACTION_MEMORY_MODE=embed_only keeps all memories searchable through embeddings without graph/entity extraction.
• EXTRACTION_MEMORY_MODE=extract_durable extracts only durable memories.
• EXTRACTION_MEMORY_MODE=extract_all extracts durable memories and session context.