Prompt Sommelier: Smart LLM Selection via Browser Extension, Say Goodbye to Token Waste

Prompt Sommelier is a browser extension that uses local ONNX inference to analyze user prompts, intelligently recommends the most suitable LLM tier, helps users make optimal choices before calling cloud-based large models, saves token costs, and protects privacy. The project is developed and maintained by Developer-Mike-Collins, with source code hosted on GitHub, and was released on June 5, 2026.

With the explosive growth of the LLM ecosystem, users face the problem of model selection: current LLMs are divided into three tiers—lightweight local models (e.g., Phi-4, Gemma 2B), mid-tier cloud models (e.g., GPT-3.5, Claude 3 Haiku), and flagship large models (e.g., GPT-4o, Claude 3.5 Sonnet). Many users are accustomed to using the strongest model for all tasks, leading to a lot of unnecessary token costs, while many daily tasks can be completed with smaller, cheaper models.

Prompt Sommelier is an innovative browser extension with the core concept of "smart matching". Its core features include: 1. Local ONNX inference—sensitive prompts are not sent to external servers; 2. Real-time analysis of prompt complexity, task type, and required capabilities; 3. Intelligent recommendation of LLM tiers; 4. Zero cloud dependency, fully offline operation (except for installation), no API key required.

The technical architecture is based on edge AI: 1. Uses ONNX Runtime for Web, supporting cross-platform compatibility, performance optimization, and size control; 2. The prompt classification model follows the path of text embedding → task classification → complexity evaluation → model matching (specific architecture not disclosed).

1. Cost optimization: Pre-screens tasks—recommends local/cheap models for simple tasks, and only uses flagship models for complex tasks; 2. Privacy protection: Analyzes sensitive prompts locally to decide whether to use cloud models; 3. Response speed: Local models have no network latency, providing a better experience for simple tasks.

Scenario 1: Daily business emails → low complexity → recommend local lightweight models or GPT-3.5 level; Scenario 2: Complex code review → high complexity → recommend GPT-4 or Claude 3.5 Sonnet level; Scenario 3: Creative writing → medium-high complexity → recommend mid-tier or above models.

Current limitations: Model coverage depends on the quality of training data, task boundaries are vague, and continuous updates are needed to adapt to new models. Future directions: Personalized learning to optimize recommendations, multi-dimensional evaluation (cost/latency/privacy), integration with mainstream LLM APIs, and enterprise-level features (team policies/cost tracking/compliance auditing).

Prompt Sommelier embodies the refined and intelligent development direction of LLM applications, and is an important tool for optimizing cost and efficiency. As edge AI technology matures, more such innovations will emerge to help users use AI more intelligently.