Query your canonical model
Learn how a query layer resolves a request against one canonical graph - lexical BM25 and dense HNSW vectors fused by Reciprocal Rank Fusion, a graph-aware rerank and ranking signals, and a bounded, read-only profile agents call directly. Cogitave Query is the reference implementation you'll trace.
Units5
Duration25 min
Levelintermediate
By the end of this module, you'll be able to:
- Explain why a query layer fuses lexical, dense, and graph signals instead of relying on any one retriever alone.
- Trace a query through the pipeline: BM25 recall, HNSW dense recall, Reciprocal Rank Fusion, and graph-aware rerank.
- State why Reciprocal Rank Fusion was chosen over a weighted linear combination of retriever scores.
- Read a bounded, read-only graph-query contract well enough to explain what makes it safe for an agent to call directly - using Cogitave Query's
query_graphprofile as the worked example.
Prerequisites
- Familiarity with a canonical property graph - the node/edge model every query resolves against - is assumed. Cogitave's Core is the worked example throughout this path.
- Recommended: The one-model architecture, the first module of this path, which covers that graph substrate directly.
Units
- 01Introduction4 min
- 02How a query layer resolves a request7 min
- 03Why hybrid retrieval, and what the graph adds7 min
- 04Knowledge check4 min
- 05Summary3 min
Related
- The one-model architecture Explain why an AI-native organization runs on one canonical property graph that humans and agents both query, what its node/edge model and identity scheme look like, and why a labeled property graph is the right substrate. Cogitave's Core is the concrete example.
- A native MCP interface to the canonical model Read a canonical model's native MCP interface as the source of truth for how humans and agents query one org's estate - why the surface is native rather than adapted, the protocol contract it commits to, and the tools and resources you call. Cogitave's Core is the worked example.
- Project products into the canonical model See how content projects into one canonical model as typed, content-addressed nodes rather than separate silos - Cogitave's Core is the worked example - and how the fact registry's cite-not-restate rule and fact-drift scanner keep one owner per fact across your estate.