Package org.mockserver.llm.client

Interface LlmClient

All Known Implementing Classes:
AbstractLlmClient, AnthropicLlmClient, AzureOpenAiLlmClient, BedrockLlmClient, GeminiLlmClient, OllamaLlmClient, OpenAiLlmClient, OpenAiResponsesLlmClient
public interface LlmClient
Per-provider knowledge for calling a real LLM as a client. Mirrors ProviderCodec (which goes the other way — decoding inbound requests / encoding mock responses): an LlmClient builds an outbound request and parses the inbound response into a Completion.

Deliberately split from transport: implementations are pure functions of their inputs (no network, no shared state), so they unit-test offline. The actual HTTP call, timeouts, fail-closed handling, and caching live in LlmCompletionService. Adding a provider = implement this interface and register it in LlmClientRegistry — the same one-line story as codecs.

  • Method Details

    • provider

      Provider provider()
      The provider this client handles — the registry key.

    • buildCompletionRequest

      HttpRequest buildCompletionRequest(LlmBackend backend, ParsedConversation prompt)
      Build the outbound completion request (URL path, auth headers, request body) for the given backend and prompt. The returned request carries the socket address resolved from the backend's base URL (or the provider default). Implementations pin temperature=0 and a seed where the provider supports it, for reproducibility.

    • parseCompletionResponse

      Completion parseCompletionResponse(HttpResponse response)
      Parse a successful provider response body into a Completion. Implementations should tolerate missing optional fields and never throw for a well-formed-but-sparse response; malformed bodies may throw and are handled (fail-closed) by LlmCompletionService.