Plan the request
The selected profile, filters, source boundaries, and caller identity determine which retrieval channels may run.
CogLake Evidence Fusion
Retrieval that returns evidence, not just similarity.
CogLake does not rely on a single similarity score. It plans each request, combines complementary retrieval channels, verifies access, and returns the evidence needed to understand every result.
QueryWhat changed in the renewal?
LexicalExact terms
SemanticMeaning
GraphRelations
Evidence FusionRank, verify, explain
3 verified sourcesFile IDs, citations, links, and access decision attached
A staged decision system
Each result has to survive the complete pipeline.
Retrieval breadth and ranking quality matter, but enterprise answers also require identity, permission, provenance, and reproducibility.
Gather candidates
Lexical, semantic, metadata, and graph channels retrieve complementary evidence without flattening their signals.
Fuse and rerank
Rank fusion, reranking, and bounded graph expansion turn independent candidates into one evidence set.
Verify before delivery
Permissions, stable file identity, citations, and original links are checked and attached to the response.
Hybrid retrieval pipeline
Different questions need different evidence.
Exact identifiers, semantic meaning, metadata, and linked context are kept as distinct signals, then combined through a controlled ranking and verification path.
Question
LexicalExact terms
DenseSemantic meaning
GraphLinked context
CogLakeEvidence FusionWeighted RRF + rerank
Verified evidenceCitations + ACL check
Explicit search profiles
Quality controls are selectable, observable, and repeatable.
Frontend, REST, and MCP callers can use the same named behavior. Administrators can inspect which channels ran instead of relying on an opaque search mode.
Search profile
Behavior is explicit for every requestProfileLexicalSemanticGraphBest for
ExactHighLowOffIDs, clauses, codes
BalancedMediumHighSelectiveEveryday knowledge work
ResearchMediumHighExpandedMulti-document investigation
Evidence and validation
Why a result appeared remains inspectable.
Every result can carry the matching signals, stable source identity, original location, permission decision, and graph path used to reach related context.
Keyword, semantic, metadata, rerank, and graph contributions stay distinguishable.
Caller identity and source ACL decisions are enforced before results leave the platform.
Stable file IDs, original links, citations, and bounded relation paths keep context traceable.
Retrieval benchmarks, ACL isolation, update consistency, and multimodal needles protect releases.
Start with your real knowledge
Evaluate search against your real questions.
Use representative documents, permission boundaries, and known answers to see how Evidence Fusion behaves in your environment.
Book a walkthrough