Summary

You now know how a query layer resolves a request against one canonical graph - with Cogitave Query as the worked example.

In this module, you:

• Named the three signals Cogitave Query fuses - lexical (BM25 over immutable segments), dense (HNSW over per-node embeddings), and graph structure - and why none of them alone is enough for technical or agent queries.
• Traced the fusion step: Reciprocal Rank Fusion, chosen over a weighted linear blend of scores because rank-based fusion sidesteps reconciling BM25's unbounded scores with cosine's bounded range, and lets retrievers change without recalibration.
• Saw what the graph-aware rerank adds once lexical and dense are already fused - a Personalized PageRank structural prior, plus ranking signals for moniker freshness, audience, prerequisite proximity, level fit, and authority - and that facets are graph traversals, not a second taxonomy to maintain.
• Read the bounded query_graph contract that lets an agent pattern-match the graph directly: read-only, allowlisted to the 9 edge labels, depth- and row-capped, timed out, and parameterized only - a hard, server-enforced contract rather than advisory guidance.

Next steps

• A native MCP interface to the canonical model - the next module in this path, which takes the surfaces this one only named - the native MCP tools and cogitave:// resources - and teaches the interface itself.
• The Cogitave Query doc, in the estate's core repository, is worth a reread directly; this module followed its section order.
• ADR-0002 - Hybrid retrieval is the decision record behind the fusion choice, with the options that were rejected and why.