Query intent and semantic routing
Classify a request by information lifecycle or domain, then route it to durable knowledge, project state, live telemetry, or another controlled source.
Technical note
Intent graphs and goal orchestration in AI systems.
Intent graphs model goals, dependencies, routes, and state transitions. They connect a raw request to the information sources, capabilities, and execution path an AI system should use.
Core distinction
Knowledge graphs describe what is connected. Intent graphs describe what should happen next.
A conventional knowledge graph represents entities and factual relationships. An intent graph represents motivations, sequences, dependencies, capability routes, and state transitions.
In practice, the two can work together. The intent layer decides which knowledge domain, live system, project state, or agent capability is relevant to the current goal.
Common graph objects
- Goals and subgoals
- Intent categories and trust boundaries
- Task dependencies and execution states
- Agent skills and reusable workflows
- Transitions caused by evidence or feedback
Four useful forms
Intent graphs appear at different layers of an AI system
The word describes a family of graph structures rather than one standardized implementation.
Task-dependency planning
Decompose a goal into subtasks whose edges represent prerequisites, data dependencies, or execution order.
Skill composition
Map atomic capabilities, reusable workflows, and higher-level orchestrators so an agent can select an appropriate execution path.
Commercial and search journeys
Map the questions, decisions, and evidence needs that connect an initial problem to later buyer or research actions.
MainFrame example
Route by intent before retrieval.
In the MainFrame architecture, query intent is used to distinguish durable knowledge from project coordination and volatile live state. This reduces the risk of answering a current-status question from a static knowledge index or leaking unprocessed inbox material into a general retrieval path.
Intent markers can also be embedded into query and document representations so retrieval aligns along functional as well as topical dimensions.
User request
→Intent classifier
→Trust-boundary router
→Knowledge, project, or live source
Design cautions
A graph is useful only when its states and evidence are real
Intent architecture can become decorative if edges do not correspond to executable routes, controlled transitions, or observed feedback.
Do not infer certainty from topology
A path through a graph does not prove that the selected source is authoritative or that a subtask completed correctly.
Keep volatile state separate
Current status, durable facts, and working hypotheses should not silently share the same retrieval and trust assumptions.
Expose unresolved routes
The system needs an explicit state for ambiguous intent, unavailable tools, missing evidence, or blocked dependencies.
Record transitions
For consequential work, retain why the system selected a route, what evidence changed state, and which component made the decision.
Researching an AI-system architecture or evidence workflow?
The research sprint can map the current landscape, competing architectural interpretations, operational risks, and the evidence needed to choose a direction.