genai-refinement-advisor

genai-refinement-advisor is a decision-support tool for GenAI delivery. Given a task (a labelled dataset) and a set of candidate configurations (foundation model x refinement level), it measures quality, cost, and latency for each, computes a cost-performance Pareto frontier, and recommends the cheapest configuration that clears a quality threshold you set. Refinement escalates through the usual ladder: zero-shot prompt -> few-shot prompt -> RAG -> agent. The tool directly supports the core consulting question: find the appropriate level of refinement (semantic search, RAG, agents, fine-tuning) to reach the value-delivery threshold in the most cost-effective way. It includes a deterministic offline mock provider, so the CLI and full test suite run with zero network access. The methodology (grid benchmark -> quality/cost/latency -> Pareto frontier -> threshold-aware recommendation) is real and provider-agnostic; the mock's quality numbers are a stand-in you replace with a real provider on your own data. Fine-tuning is named as the next rung of the ladder but is not implemented here.