A lightweight, fast, and fully deterministic alternative to traditional intent-classification systems.
Built with React + Vite + Tailwind on the frontend and a Python-based semantic inference engine on the backend.
This system routes user queries through a graph of intents using sentence embeddings, adaptive confidence logic, and retrieval-based fallbacks—no classifiers, training loops, or model deployments required.
Resolves user queries by traversing a DAG-structured intent graph with path-level scoring instead of a single classifier.
Automatically adjusts sensitivity based on query length and phrasing, improving routing stability on ambiguous inputs.
When a query doesn’t clearly match any intent, the system fetches semantically similar candidates and recovers gracefully.
Maintains conversational context so follow-up questions like “same as before” or “for that” route correctly without repeating selections.
Intents are defined in JSON and automatically converted into a navigable graph.
Updates apply instantly—no retraining or redeployment required.
Runs entirely on CPU and maintains sub-15ms routing latency thanks to caching and optimized traversal.
- Frontend: React + Vite + Tailwind interface for entering queries and testing the routing behavior.
- Backend: Python engine using sentence-transformer embeddings and deterministic traversal logic.
- Intent Graph: JSON-defined structure supporting multi-parent nodes, examples, responses, and metadata.
Most NLU systems rely on classifiers or fine-tuned models, which brings problems like:
- retraining loops
- model drift
- slow iteration cycles
- low explainability
This project avoids all of that by using semantic similarity, graph traversal, and context tracking to produce stable and predictable routing—even as intents change.