Edge-first intent classifier that routes queries to the right data sources automatically. The AI traffic controller that makes everything faster and smarter.
The intent classifier runs on Cloudflare Workers AI at the edge, classifying user queries in single-digit milliseconds before the main AI model even starts processing. This pre-classification step determines which data sources to query — financial data for spend questions, task data for project questions, knowledge base for process questions — so the AI retrieval system fetches exactly the right context.
The classifier recognizes multiple intent types: financial queries, project status, search requests, process questions, client information, time tracking queries, and general conversation. Each intent type has a tailored scoring profile that adjusts the weights for semantic similarity, recency, importance, and entity matching in the retrieval pipeline. A financial query prioritizes recency; a process query prioritizes semantic accuracy.
When a LoRA adapter trained on your agency's data is available, the classifier uses it as the primary classification model. Agency-specific language patterns, custom terminology, and domain jargon are understood natively rather than requiring generic model interpretation. The system falls back to the base model when no adapter is available, ensuring consistent functionality.
Classification accuracy improves over time through the feedback loop. When users indicate that a response was not helpful, the system logs the original intent classification for review. Patterns of misclassification inform retraining priorities and LoRA adapter updates. The goal is a classifier that understands your agency's language as well as your team does.