@ai-sdk/openai-compatible 2.0.15 → 2.0.17

package/src/completion/openai-compatible-completion-language-model.ts

@@ -0,0 +1,390 @@
+import {
+  APICallError,
+  LanguageModelV3,
+  LanguageModelV3CallOptions,
+  LanguageModelV3Content,
+  LanguageModelV3FinishReason,
+  LanguageModelV3GenerateResult,
+  LanguageModelV3StreamPart,
+  LanguageModelV3StreamResult,
+  SharedV3Warning,
+} from '@ai-sdk/provider';
+import {
+  combineHeaders,
+  createEventSourceResponseHandler,
+  createJsonErrorResponseHandler,
+  createJsonResponseHandler,
+  FetchFunction,
+  parseProviderOptions,
+  ParseResult,
+  postJsonToApi,
+  ResponseHandler,
+} from '@ai-sdk/provider-utils';
+import { z } from 'zod/v4';
+import {
+  defaultOpenAICompatibleErrorStructure,
+  ProviderErrorStructure,
+} from '../openai-compatible-error';
+import { convertOpenAICompatibleCompletionUsage } from './convert-openai-compatible-completion-usage';
+import { convertToOpenAICompatibleCompletionPrompt } from './convert-to-openai-compatible-completion-prompt';
+import { getResponseMetadata } from './get-response-metadata';
+import { mapOpenAICompatibleFinishReason } from './map-openai-compatible-finish-reason';
+import {
+  OpenAICompatibleCompletionModelId,
+  openaiCompatibleCompletionProviderOptions,
+} from './openai-compatible-completion-options';
+
+type OpenAICompatibleCompletionConfig = {
+  provider: string;
+  includeUsage?: boolean;
+  headers: () => Record<string, string | undefined>;
+  url: (options: { modelId: string; path: string }) => string;
+  fetch?: FetchFunction;
+  errorStructure?: ProviderErrorStructure<any>;
+
+  /**
+   * The supported URLs for the model.
+   */
+  supportedUrls?: () => LanguageModelV3['supportedUrls'];
+};
+
+export class OpenAICompatibleCompletionLanguageModel
+  implements LanguageModelV3
+{
+  readonly specificationVersion = 'v3';
+
+  readonly modelId: OpenAICompatibleCompletionModelId;
+  private readonly config: OpenAICompatibleCompletionConfig;
+  private readonly failedResponseHandler: ResponseHandler<APICallError>;
+  private readonly chunkSchema; // type inferred via constructor
+
+  constructor(
+    modelId: OpenAICompatibleCompletionModelId,
+    config: OpenAICompatibleCompletionConfig,
+  ) {
+    this.modelId = modelId;
+    this.config = config;
+
+    // initialize error handling:
+    const errorStructure =
+      config.errorStructure ?? defaultOpenAICompatibleErrorStructure;
+    this.chunkSchema = createOpenAICompatibleCompletionChunkSchema(
+      errorStructure.errorSchema,
+    );
+    this.failedResponseHandler = createJsonErrorResponseHandler(errorStructure);
+  }
+
+  get provider(): string {
+    return this.config.provider;
+  }
+
+  private get providerOptionsName(): string {
+    return this.config.provider.split('.')[0].trim();
+  }
+
+  get supportedUrls() {
+    return this.config.supportedUrls?.() ?? {};
+  }
+
+  private async getArgs({
+    prompt,
+    maxOutputTokens,
+    temperature,
+    topP,
+    topK,
+    frequencyPenalty,
+    presencePenalty,
+    stopSequences: userStopSequences,
+    responseFormat,
+    seed,
+    providerOptions,
+    tools,
+    toolChoice,
+  }: LanguageModelV3CallOptions) {
+    const warnings: SharedV3Warning[] = [];
+
+    // Parse provider options
+    const completionOptions =
+      (await parseProviderOptions({
+        provider: this.providerOptionsName,
+        providerOptions,
+        schema: openaiCompatibleCompletionProviderOptions,
+      })) ?? {};
+
+    if (topK != null) {
+      warnings.push({ type: 'unsupported', feature: 'topK' });
+    }
+
+    if (tools?.length) {
+      warnings.push({ type: 'unsupported', feature: 'tools' });
+    }
+
+    if (toolChoice != null) {
+      warnings.push({ type: 'unsupported', feature: 'toolChoice' });
+    }
+
+    if (responseFormat != null && responseFormat.type !== 'text') {
+      warnings.push({
+        type: 'unsupported',
+        feature: 'responseFormat',
+        details: 'JSON response format is not supported.',
+      });
+    }
+
+    const { prompt: completionPrompt, stopSequences } =
+      convertToOpenAICompatibleCompletionPrompt({ prompt });
+
+    const stop = [...(stopSequences ?? []), ...(userStopSequences ?? [])];
+
+    return {
+      args: {
+        // model id:
+        model: this.modelId,
+
+        // model specific settings:
+        echo: completionOptions.echo,
+        logit_bias: completionOptions.logitBias,
+        suffix: completionOptions.suffix,
+        user: completionOptions.user,
+
+        // standardized settings:
+        max_tokens: maxOutputTokens,
+        temperature,
+        top_p: topP,
+        frequency_penalty: frequencyPenalty,
+        presence_penalty: presencePenalty,
+        seed,
+        ...providerOptions?.[this.providerOptionsName],
+
+        // prompt:
+        prompt: completionPrompt,
+
+        // stop sequences:
+        stop: stop.length > 0 ? stop : undefined,
+      },
+      warnings,
+    };
+  }
+
+  async doGenerate(
+    options: LanguageModelV3CallOptions,
+  ): Promise<LanguageModelV3GenerateResult> {
+    const { args, warnings } = await this.getArgs(options);
+
+    const {
+      responseHeaders,
+      value: response,
+      rawValue: rawResponse,
+    } = await postJsonToApi({
+      url: this.config.url({
+        path: '/completions',
+        modelId: this.modelId,
+      }),
+      headers: combineHeaders(this.config.headers(), options.headers),
+      body: args,
+      failedResponseHandler: this.failedResponseHandler,
+      successfulResponseHandler: createJsonResponseHandler(
+        openaiCompatibleCompletionResponseSchema,
+      ),
+      abortSignal: options.abortSignal,
+      fetch: this.config.fetch,
+    });
+
+    const choice = response.choices[0];
+    const content: Array<LanguageModelV3Content> = [];
+
+    // text content:
+    if (choice.text != null && choice.text.length > 0) {
+      content.push({ type: 'text', text: choice.text });
+    }
+
+    return {
+      content,
+      usage: convertOpenAICompatibleCompletionUsage(response.usage),
+      finishReason: {
+        unified: mapOpenAICompatibleFinishReason(choice.finish_reason),
+        raw: choice.finish_reason,
+      },
+      request: { body: args },
+      response: {
+        ...getResponseMetadata(response),
+        headers: responseHeaders,
+        body: rawResponse,
+      },
+      warnings,
+    };
+  }
+
+  async doStream(
+    options: LanguageModelV3CallOptions,
+  ): Promise<LanguageModelV3StreamResult> {
+    const { args, warnings } = await this.getArgs(options);
+
+    const body = {
+      ...args,
+      stream: true,
+
+      // only include stream_options when in strict compatibility mode:
+      stream_options: this.config.includeUsage
+        ? { include_usage: true }
+        : undefined,
+    };
+
+    const { responseHeaders, value: response } = await postJsonToApi({
+      url: this.config.url({
+        path: '/completions',
+        modelId: this.modelId,
+      }),
+      headers: combineHeaders(this.config.headers(), options.headers),
+      body,
+      failedResponseHandler: this.failedResponseHandler,
+      successfulResponseHandler: createEventSourceResponseHandler(
+        this.chunkSchema,
+      ),
+      abortSignal: options.abortSignal,
+      fetch: this.config.fetch,
+    });
+
+    let finishReason: LanguageModelV3FinishReason = {
+      unified: 'other',
+      raw: undefined,
+    };
+    let usage:
+      | {
+          prompt_tokens: number | undefined;
+          completion_tokens: number | undefined;
+          total_tokens: number | undefined;
+        }
+      | undefined = undefined;
+    let isFirstChunk = true;
+
+    return {
+      stream: response.pipeThrough(
+        new TransformStream<
+          ParseResult<z.infer<typeof this.chunkSchema>>,
+          LanguageModelV3StreamPart
+        >({
+          start(controller) {
+            controller.enqueue({ type: 'stream-start', warnings });
+          },
+
+          transform(chunk, controller) {
+            if (options.includeRawChunks) {
+              controller.enqueue({ type: 'raw', rawValue: chunk.rawValue });
+            }
+
+            // handle failed chunk parsing / validation:
+            if (!chunk.success) {
+              finishReason = { unified: 'error', raw: undefined };
+              controller.enqueue({ type: 'error', error: chunk.error });
+              return;
+            }
+
+            const value = chunk.value;
+
+            // handle error chunks:
+            if ('error' in value) {
+              finishReason = { unified: 'error', raw: undefined };
+              controller.enqueue({ type: 'error', error: value.error });
+              return;
+            }
+
+            if (isFirstChunk) {
+              isFirstChunk = false;
+
+              controller.enqueue({
+                type: 'response-metadata',
+                ...getResponseMetadata(value),
+              });
+
+              controller.enqueue({
+                type: 'text-start',
+                id: '0',
+              });
+            }
+
+            if (value.usage != null) {
+              usage = value.usage;
+            }
+
+            const choice = value.choices[0];
+
+            if (choice?.finish_reason != null) {
+              finishReason = {
+                unified: mapOpenAICompatibleFinishReason(choice.finish_reason),
+                raw: choice.finish_reason ?? undefined,
+              };
+            }
+
+            if (choice?.text != null) {
+              controller.enqueue({
+                type: 'text-delta',
+                id: '0',
+                delta: choice.text,
+              });
+            }
+          },
+
+          flush(controller) {
+            if (!isFirstChunk) {
+              controller.enqueue({ type: 'text-end', id: '0' });
+            }
+
+            controller.enqueue({
+              type: 'finish',
+              finishReason,
+              usage: convertOpenAICompatibleCompletionUsage(usage),
+            });
+          },
+        }),
+      ),
+      request: { body },
+      response: { headers: responseHeaders },
+    };
+  }
+}
+
+const usageSchema = z.object({
+  prompt_tokens: z.number(),
+  completion_tokens: z.number(),
+  total_tokens: z.number(),
+});
+
+// limited version of the schema, focussed on what is needed for the implementation
+// this approach limits breakages when the API changes and increases efficiency
+const openaiCompatibleCompletionResponseSchema = z.object({
+  id: z.string().nullish(),
+  created: z.number().nullish(),
+  model: z.string().nullish(),
+  choices: z.array(
+    z.object({
+      text: z.string(),
+      finish_reason: z.string(),
+    }),
+  ),
+  usage: usageSchema.nullish(),
+});
+
+// limited version of the schema, focussed on what is needed for the implementation
+// this approach limits breakages when the API changes and increases efficiency
+const createOpenAICompatibleCompletionChunkSchema = <
+  ERROR_SCHEMA extends z.core.$ZodType,
+>(
+  errorSchema: ERROR_SCHEMA,
+) =>
+  z.union([
+    z.object({
+      id: z.string().nullish(),
+      created: z.number().nullish(),
+      model: z.string().nullish(),
+      choices: z.array(
+        z.object({
+          text: z.string(),
+          finish_reason: z.string().nullish(),
+          index: z.number(),
+        }),
+      ),
+      usage: usageSchema.nullish(),
+    }),
+    errorSchema,
+  ]);

package/src/completion/openai-compatible-completion-options.ts

@@ -0,0 +1,33 @@
+import { z } from 'zod/v4';
+
+export type OpenAICompatibleCompletionModelId = string;
+
+export const openaiCompatibleCompletionProviderOptions = z.object({
+  /**
+   * Echo back the prompt in addition to the completion.
+   */
+  echo: z.boolean().optional(),
+
+  /**
+   * Modify the likelihood of specified tokens appearing in the completion.
+   *
+   * Accepts a JSON object that maps tokens (specified by their token ID in
+   * the GPT tokenizer) to an associated bias value from -100 to 100.
+   */
+  logitBias: z.record(z.string(), z.number()).optional(),
+
+  /**
+   * The suffix that comes after a completion of inserted text.
+   */
+  suffix: z.string().optional(),
+
+  /**
+   * A unique identifier representing your end-user, which can help providers to
+   * monitor and detect abuse.
+   */
+  user: z.string().optional(),
+});
+
+export type OpenAICompatibleCompletionProviderOptions = z.infer<
+  typeof openaiCompatibleCompletionProviderOptions
+>;

package/src/embedding/openai-compatible-embedding-model.test.ts

@@ -0,0 +1,171 @@
+import { describe, it, expect } from 'vitest';
+import { EmbeddingModelV3Embedding } from '@ai-sdk/provider';
+import { createTestServer } from '@ai-sdk/test-server/with-vitest';
+import { createOpenAICompatible } from '../openai-compatible-provider';
+
+const dummyEmbeddings = [
+  [0.1, 0.2, 0.3, 0.4, 0.5],
+  [0.6, 0.7, 0.8, 0.9, 1.0],
+];
+const testValues = ['sunny day at the beach', 'rainy day in the city'];
+
+const provider = createOpenAICompatible({
+  baseURL: 'https://my.api.com/v1/',
+  name: 'test-provider',
+  headers: {
+    Authorization: `Bearer test-api-key`,
+  },
+});
+const model = provider.embeddingModel('text-embedding-3-large');
+
+const server = createTestServer({
+  'https://my.api.com/v1/embeddings': {},
+});
+
+describe('doEmbed', () => {
+  function prepareJsonResponse({
+    embeddings = dummyEmbeddings,
+    usage = { prompt_tokens: 8, total_tokens: 8 },
+    headers,
+  }: {
+    embeddings?: EmbeddingModelV3Embedding[];
+    usage?: { prompt_tokens: number; total_tokens: number };
+    headers?: Record<string, string>;
+  } = {}) {
+    server.urls['https://my.api.com/v1/embeddings'].response = {
+      type: 'json-value',
+      headers,
+      body: {
+        object: 'list',
+        data: embeddings.map((embedding, i) => ({
+          object: 'embedding',
+          index: i,
+          embedding,
+        })),
+        model: 'text-embedding-3-large',
+        usage,
+      },
+    };
+  }
+
+  it('should extract embedding', async () => {
+    prepareJsonResponse();
+
+    const { embeddings } = await model.doEmbed({ values: testValues });
+
+    expect(embeddings).toStrictEqual(dummyEmbeddings);
+  });
+
+  it('should expose the raw response headers', async () => {
+    prepareJsonResponse({
+      headers: { 'test-header': 'test-value' },
+    });
+
+    const { response } = await model.doEmbed({ values: testValues });
+
+    expect(response?.headers).toStrictEqual({
+      // default headers:
+      'content-length': '236',
+      'content-type': 'application/json',
+
+      // custom header
+      'test-header': 'test-value',
+    });
+  });
+
+  it('should extract usage', async () => {
+    prepareJsonResponse({
+      usage: { prompt_tokens: 20, total_tokens: 20 },
+    });
+
+    const { usage } = await model.doEmbed({ values: testValues });
+
+    expect(usage).toStrictEqual({ tokens: 20 });
+  });
+
+  it('should pass the model and the values', async () => {
+    prepareJsonResponse();
+
+    await model.doEmbed({ values: testValues });
+
+    expect(await server.calls[0].requestBodyJson).toStrictEqual({
+      model: 'text-embedding-3-large',
+      input: testValues,
+      encoding_format: 'float',
+    });
+  });
+
+  it('should pass the dimensions setting', async () => {
+    prepareJsonResponse();
+
+    await provider.embeddingModel('text-embedding-3-large').doEmbed({
+      values: testValues,
+      providerOptions: {
+        openaiCompatible: {
+          dimensions: 64,
+        },
+      },
+    });
+
+    expect(await server.calls[0].requestBodyJson).toStrictEqual({
+      model: 'text-embedding-3-large',
+      input: testValues,
+      encoding_format: 'float',
+      dimensions: 64,
+    });
+  });
+
+  it('should pass settings with deprecated openai-compatible key and emit warning', async () => {
+    prepareJsonResponse();
+
+    const result = await provider
+      .embeddingModel('text-embedding-3-large')
+      .doEmbed({
+        values: testValues,
+        providerOptions: {
+          'openai-compatible': {
+            dimensions: 64,
+          },
+        },
+      });
+
+    expect(await server.calls[0].requestBodyJson).toStrictEqual({
+      model: 'text-embedding-3-large',
+      input: testValues,
+      encoding_format: 'float',
+      dimensions: 64,
+    });
+
+    expect(result.warnings).toContainEqual({
+      type: 'other',
+      message: `The 'openai-compatible' key in providerOptions is deprecated. Use 'openaiCompatible' instead.`,
+    });
+  });
+
+  it('should pass headers', async () => {
+    prepareJsonResponse();
+
+    const provider = createOpenAICompatible({
+      baseURL: 'https://my.api.com/v1/',
+      name: 'test-provider',
+      headers: {
+        Authorization: `Bearer test-api-key`,
+        'Custom-Provider-Header': 'provider-header-value',
+      },
+    });
+
+    await provider.embeddingModel('text-embedding-3-large').doEmbed({
+      values: testValues,
+      headers: {
+        'Custom-Request-Header': 'request-header-value',
+      },
+    });
+
+    expect(server.calls[0].requestHeaders).toStrictEqual({
+      authorization: 'Bearer test-api-key',
+      'content-type': 'application/json',
+      'custom-provider-header': 'provider-header-value',
+      'custom-request-header': 'request-header-value',
+    });
+  });
+});