VectorTokenizer

VectorTokenizer: Semantic Hybrid-Routing for AI Economics Enterprises face an architectural dilemma: routing all tasks to premium LLMs burns API budgets, while relying solely on local models degrades complex reasoning. To conquer the Track 1 challenge of maintaining 80 percent accuracy with minimal token spend, we built VectorTokenizer, a headless, containerized hybrid-routing gateway. The Semantic Routing Advantage Most baselines use brittle regex or keyword matching, which shatter on unseen prompt variations and cause devastating misclassifications. VectorTokenizer replaces these fragile rules with sub-5ms semantic intent classification. We pre-load the all-MiniLM-L6-v2 embedding model into RAM, vectorizing incoming tasks and mapping them to eight core domains using Cosine Similarity. The Escalation Ladder Tasks route through a strict deterministic pipeline. First, Zero-Cost Compute. Routine workloads like Factual, Sentiment, and Summarization hit a 4-bit quantized Phi-3-mini executing entirely inside the container via llama-cpp-python, costing exactly zero cloud tokens. Second, Cloud Escalation. High-complexity domains like Code Generation and Logic safely escalate to the Fireworks AI API. Production-Grade MLOps To beat the strict 60-second boot and 10GB image limits, we utilized a Docker Bake-In strategy. Model weights are downloaded during the build phase. The resulting 5.2GB image completely eliminates network cold starts, booting into RAM in under 5 seconds. Deterministic Guardrails The agent dynamically parses ALLOWED_MODELS to filter out trap models, locks temperature at 0.0 for strict accuracy, and enforces robust retry loops to guarantee a successful run.