Skip to content

Search and Retrieval

This page keeps the historical /features/hybrid-search slug, but the current retrieval model should be described as graph-led retrieval.

Valter no longer treats the knowledge graph as a small optional score boost layered on top of a search-first pipeline. The current model uses the graph as a primary path for discovery and explanation, while lexical and semantic search remain complementary signals and fallback mechanisms.

In practice, the retrieval stack now combines:

  • graph discovery and traversal for structure-aware relevance
  • lexical and semantic retrieval for parallel evidence gathering
  • reranking and fusion where helpful
  • explainability outputs that describe why a decision surfaced
  • graceful degradation when graph paths are unavailable

The difference is not just wording. It changes how consumers should think about Valter:

  • relevance is not only a weighted search score;
  • retrieval can be explained through entities, paths, and precedent structure;
  • search remains important, but not as the sole center of the system.
EndpointPurpose
POST /v1/retrieveRetrieve jurisprudence through the current graph-led pipeline
POST /v1/similar_casesFind structurally and semantically similar decisions
POST /v1/search/featuresFilter over extracted legal metadata
POST /v1/factual/dual-searchCompare factual and thesis-oriented retrieval views
  • Compatibility knobs and low-level weights may still exist in the runtime during migration.
  • The conceptual model should still be described as graph-led retrieval.
  • If you need strict request and response details for a client, prefer the current runtime schema or generated API surface over manually copied parameter inventories.