Introduction

An AI-native org runs on one canonical model: a single typed property graph shared by every product, standard, and agent in the estate (The one-model architecture covers that graph directly). Holding the graph is only half the story - you also need one way to ask it something. That layer is the query interface over the model, and Cogitave Query is the reference implementation this module traces.

One query layer over the canonical graph. Same model, same retrieval for humans (API/UI) and agents (MCP). Retrieval is hybrid: lexical recall, semantic recall, and graph structure are three signals fused into one ranking - none alone is enough.

That single sentence from Cogitave Query captures the pattern this module teaches. With one query layer, there is no separate "agent view" of the estate and a different "human view" - in Cogitave's estate, docs_search over MCP and the human-facing search box call the same pipeline and get the same ranking. This module teaches you to reason about what such a pipeline does to a query, so you can design your own - not to reimplement Cogitave's.

Why fuse three signals instead of picking the best one

Lexical retrieval (BM25) and dense-vector retrieval each fail on inputs the other handles well: BM25 misses paraphrase and synonymy, dense retrieval misses the rare exact tokens - identifiers, error codes, API symbols, version strings - that dominate technical and agent queries. A query layer built over a canonical graph can add a third signal neither text index can see at all: the graph itself, with its prerequisite edges, cross-references, and authority structure. Cogitave Query fuses all three.

Read this as a Day 0 spec

Cogitave's Core is a Day 0 design - the canonical model, its query layer, and its native MCP surface are specified and built against, not a black box you take on faith. This module follows that same posture: every mechanism it names - BM25 segments, HNSW vectors, Reciprocal Rank Fusion, the bounded query_graph profile - is stated exactly as Cogitave Query and ADR-0002 specify it, with a link back to the source.

What you will get from this module

By the end, you will be able to trace a query through the pipeline end to end, explain why rank-based fusion was chosen over a weighted score blend, and read the bounded graph-query contract that lets an agent call the graph directly without an unbounded write path opening underneath it.