Vendor Agnostic

Normalized invocation interface across LLM providers—not output equivalence.

Problem

Each provider has incompatible APIs: OpenAI uses messages, Anthropic uses messages with different schemas, local models use OpenAI-compatible endpoints but break on response_format. Switching from OpenAI to Claude requires rewriting request logic, error handling, retry semantics. Estimated migration cost: 2–4 engineering weeks.

Neusnap Behavior

Adapters normalize request/response shapes across providers. Developer writes against a single interface. Neusnap translates to provider-specific API calls. Retry semantics are preserved across providers. Provider-specific rate limits are handled internally (exponential backoff, fallback on 429).

Guaranteed

• →No workflow code changes when swapping providers (config-only change)
• →Fallback across vendors works inside the same transaction
• →Retry semantics consistent across all providers (exponential backoff + jitter)
• →Authentication handled per adapter (API keys never exposed to client)

NOT Guaranteed

Example: Provider Swap

// Application code (unchanged)
await neusnap.llm({
  model: "openai/gpt-4o",
  prompt: "Explain ACID transactions"
});

// Config change only (no code modification)
{
  "routingPolicy": {
    "primary": "anthropic/claude-3-opus",  // swapped from OpenAI
    "fallback": ["openai/gpt-4o"]
  }
}

// Adapter handles translation:
// - OpenAI: POST /v1/chat/completions
// - Anthropic: POST /v1/messages
// - Developer sees identical interface

Supported Providers