A conventional retriever ranks chunks against one query. That works well for a question answered by one self-contained passage. Questions that combine facts across multiple entities and periods can need several searches and an explicit join in reasoning.
Suppose a user asks which equipment providers served projects that were delayed in two different years. One passage records a project delay; another lists the project’s provider. Neither passage alone answers the request.
Decompose the information need
Swipe or scroll sideways to read the full diagram.
Decompose the question into delay-event lookup and project-provider relationship lookup; join on stable IDs, verify dates and provenance, then assemble the answer.
Query rewriting expands aliases or clarifies ambiguous entities, improving search recall if the original words poorly match stored sources. Multi-query retrieval searches several distinct formulations, then deduplicates candidates. HyDE (Hypothetical Document Embeddings) embeds a model-generated hypothetical answer as a retrieval query; it can increase semantic recall but can also steer search toward plausible yet false assumptions. Evaluate the resulting retrieved evidence, not the hypothetical text’s fluency.
When a graph changes the workload
A knowledge graph represents entities as nodes and relations as edges, often with source references and effective dates. A graph traversal can answer relationship-heavy questions by walking project → contract → supplier → region paths. The graph still needs ingestion, entity resolution, deduplicated IDs and relation freshness.
GraphRAG approaches vary. Some use graph traversal to gather facts; others derive community summaries to answer broad corpus-level questions. These have different index-build costs and different failure modes. An enterprise graph built from large language model (LLM)-extracted relations can propagate entity-merging errors unless edges retain provenance and confidence.
| Workload | Starting approach | When to escalate |
|---|---|---|
| One passage answers question | Hybrid lexical+dense retrieval | Weak lexical recall or complex context |
| Several independent facts | Decomposed or multi-query retrieval | Dependencies and joins become repetitive |
| Connected entity relations | SQL joins or graph traversal | Many hops, ambiguous paths and evolving relationships |
| Broad cross-document themes | Summaries or graph communities | Corpus-level analysis worth index maintenance |
Evaluation must reward complete evidence
A first relevant result at rank one can produce perfect mean reciprocal rank (MRR) while a necessary second result is missing. Define the relevant evidence set for each multi-hop case, then measure whether all required hops survive retrieval, reranking and context packing.
Check join-key collisions, relation timestamps, alias resolution and missing-edge behavior. The answer should abstain when one critical relation remains unverified. Run a baseline without graph traversal first; knowledge graphs add extraction, updates and reconciliation overhead, so their advantage must be measurable.
Continue with ranking metrics and oracle stage tests.