methodology · 3 min read
Every provider, one test surface
OpenAI, Anthropic, Google, Azure, Mistral, Groq, xAI, DeepSeek, Cohere, Bedrock, self-hosted. One adapter, consistent findings.
The problem with per-provider testing
Each provider has its own SDK, its own idea of a "file", its own admin surface. Test them one at a time and you get results you cannot compare.
One surface
Model Inject normalises providers behind one adapter. A key is a key: we enumerate what it reaches and probe what it answers, and the finding looks the same whether it is OpenAI or a self-hosted Llama. OpenAI-compatible providers (Groq, xAI, DeepSeek, Mistral, Together, OpenRouter, Perplexity, Ollama) share one path; Anthropic and Google have their own.
Why it matters
Your stack is probably not single-vendor. The agent runs on one model, the embeddings on another, a fallback on a third. Testing them on one surface, with one scoring model, is the only way to see your real exposure rather than a pile of incomparable reports.