Skip to content

Querying the Graph with LLMs

Placeholder — this page needs to be written.

This page covers how LLMs navigate and query the knowledge graph in practice — what patterns work, what to avoid, and how to structure agent prompts for effective graph use.

The basic interaction pattern

  1. The LLM receives a user question.
  2. It calls search_entities to find relevant starting nodes.
  3. It calls get_relationships or traverse to explore the neighborhood.
  4. It synthesizes an answer from the returned subgraph, citing provenance.

Grounding and citation

A key advantage of the knowledge graph over pure RAG is that every claim the LLM makes can be grounded in a specific graph assertion with a provenance chain. Prompts should instruct the LLM to cite entity IDs and source documents in its answers.

Multi-hop reasoning

For questions that require following a chain of relationships (e.g. "which drugs interact with enzymes that metabolize compound X?"), the traverse tool is more efficient than repeated get_relationships calls. Set max_hops conservatively to avoid returning very large subgraphs.

Handling uncertainty

When the graph contains conflicting claims or low-confidence assertions, the LLM should surface the uncertainty rather than picking one answer. Prompt guidance: "If sources disagree, report the disagreement and cite both."

See also