Graph Hints
Graph hints automatically inject related knowledge-graph entities into the agent's system prompt. This surfaces cross-session connections the agent might not find through keyword search alone — for example, pulling in a related project, person, or decision that is semantically linked to the current query.
How It Works
- Before every agent turn,
getRelatedGraphHints()runs a graph search over the last user message - Up to 3 results are returned, each with up to 2 connected entities
- The injected block is prefixed
## Related Contextand appears in the system prompt before conversation history - Only entities with connected neighbours are included; isolated entities are skipped
Injected Format
text
## Related Context
- [project] Astra Core: Core runtime library (related to Open Astra)
- [person] Alice: Lead engineer on auth module (related to authentication)
- [decision] Use pgvector: Chosen for semantic search (related to memory design)Relationship to Graph Memory
Graph hints are powered by the knowledge graph. See Graph Memory for how entities and relationships are built and maintained.
ℹGraph hints are always on. To reduce noise, ensure your graph entities have descriptive names and that
connectedEntities are populated via the auto-extraction pipeline.