Class RelaceApi

packages/openai-adapters/src/apis/Relace.ts:39–178

// Relace only supports apply through a /v1/apply endpoint
export class RelaceApi implements BaseLlmApi {
  private apiBase = "https://instantapply.endpoint.relace.run/v1/";

  constructor(private readonly config: z.infer<typeof OpenAIConfigSchema>) {
    this.apiBase = config.apiBase ?? this.apiBase;
    if (!this.apiBase.endsWith("/")) {
      this.apiBase += "/";
    }
    this.config = config;
  }

  async chatCompletionNonStream(
    body: ChatCompletionCreateParamsNonStreaming,
    signal: AbortSignal,
  ): Promise<ChatCompletion> {
    let content = "";
    let usage: UsageInfo | undefined = undefined;

    // Convert the non-streaming params to streaming params
    const streamingBody: ChatCompletionCreateParamsStreaming = {
      ...body,
      stream: true,
    };

    for await (const chunk of this.chatCompletionStream(
      streamingBody,
      signal,
    )) {
      if (chunk.choices.length > 0) {
        content += chunk.choices[0]?.delta?.content || "";
      }
      if (chunk.usage) {
        usage = chunk.usage;
      }
    }

    return chatCompletion({
      content,
      model: body.model,
      usage,
    });
  }

  // We convert from what would be sent to OpenAI (a prediction for the existing code and a user message with the new code)
  // to Relace's format
  async *chatCompletionStream(
    body: ChatCompletionCreateParamsStreaming,
    signal: AbortSignal,
  ): AsyncGenerator<ChatCompletionChunk> {
    const headers = {
      "Content-Type": "application/json",
      Authorization: `Bearer ${this.config.apiKey}`,
    };

    const prediction = body.prediction?.content ?? "";
    const initialCode =
      typeof prediction === "string"
        ? prediction