Retrieval methodologies
Having multiple retrieval strategies isn’t just a feature list — it’s the difference between getting useful answers and getting noise. Each approach has different strengths, and the right choice depends on your data and your questions.
Graph RAG
Graph RAG is TrustGraph’s flagship retrieval mechanism, in development since 2023.
How it works: Documents are chunked and processed through knowledge extraction. An LLM identifies entities and their relationships, storing them in a context hypergraph with vector embeddings for each entity.
When a question is asked, TrustGraph extracts key concepts from the query, finds relevant entities via semantic similarity, then traverses the knowledge graph to discover related information. A cross-encoder reranker scores the relevance of discovered relationships. The result is a focused subgraph of precisely relevant knowledge — not a pile of text chunks.
When to use it:
- Questions that require understanding relationships between things
- Answers that need context from multiple documents
- Connecting disparate information across a large corpus
- When reducing hallucination is critical
The advantage: Graph RAG understands structure. “Who reports to the VP of Engineering?” requires traversing a reporting chain — vector similarity alone can’t do this.
Ontology RAG
Ontology RAG extends Graph RAG with domain-specific precision.
How it works: You Bring Your Own Ontology (BYOO) — a structured description of the entity types and relationships that matter in your domain. Think of it as a schema for human knowledge. TrustGraph uses this ontology to guide extraction, ensuring the hypergraph reflects your domain model rather than whatever the LLM happens to find.
The clever part: rather than flooding the LLM context window with the full ontology, TrustGraph uses a retrieval operation on the ontology itself to select the relevant subset for each chunk being processed.
When to use it:
- Domains with well-defined structures (legal, medical, financial, technical)
- When you need consistent, predictable extraction across documents
- When the quality of the knowledge graph matters more than speed of setup
The advantage: Precision. Instead of hoping the LLM extracts the right things, you define what “right” means.
Document RAG
Document RAG is the familiar vector similarity approach — enhanced in TrustGraph with concept extraction and full explainability.
How it works: Document chunks are embedded and stored in a vector database. At query time, TrustGraph extracts key concepts from the question (not just raw embedding), retrieves relevant chunks, and generates an answer with full provenance tracking.
TrustGraph also supports hybrid retrieval — combining BM25 keyword search with vector similarity using Reciprocal Rank Fusion — for improved recall.
When to use it:
- Broad semantic search across unstructured content
- When relationship structure isn’t the primary concern
- As a complement to Graph RAG for different types of questions
The trade-off: Document RAG is simpler to set up (no knowledge extraction step), but retrieved chunks are isolated text fragments with no understanding of how concepts relate to each other.
Choosing a strategy
| Graph RAG | Ontology RAG | Document RAG | |
|---|---|---|---|
| Best for | Relationship-rich queries | Domain-specific precision | Broad semantic search |
| Setup effort | Medium | Higher (BYOO) | Low |
| Extraction cost | LLM tokens at ingest | LLM tokens at ingest | Embeddings only |
| Relationship awareness | Yes | Yes, guided by schema | No |
| Explainability | Full | Full | Full |
In practice, many deployments use more than one strategy. Graph RAG or Ontology RAG for structured knowledge queries, Document RAG for broader searches.
Next
Explainability — how TrustGraph traces answers back to their sources.