A custom azure-openai-completion-provider for the vercel ai-sdk, provided as one file

custom-azure-openai-completion-provider.ts

import {
  type LanguageModelV1,
  type LanguageModelV1FinishReason,
  type LanguageModelV1LogProbs,
  type LanguageModelV1StreamPart,
  UnsupportedFunctionalityError,
  InvalidPromptError,
  type LanguageModelV1Prompt,
} from '@ai-sdk/provider';
import {
  type ParseResult,
  createEventSourceResponseHandler,
  createJsonResponseHandler,
  createJsonErrorResponseHandler,
  postJsonToApi,
} from '@ai-sdk/provider-utils';
import { z } from 'zod';
import {
  type OpenAICompletionModelId,
  type OpenAICompletionSettings,
} from '@ai-sdk/openai/internal';

export const openAIErrorDataSchema = z.object({
  error: z.object({
    message: z.string(),
    type: z.string(),
    param: z.any().nullable(),
    code: z.string().nullable(),
  }),
});

export type OpenAIErrorData = z.infer<typeof openAIErrorDataSchema>;

export const openaiFailedResponseHandler = createJsonErrorResponseHandler({
  errorSchema: openAIErrorDataSchema,
  errorToMessage: data => data.error.message,
});

export function mapOpenAIFinishReason(
  finishReason: string | null | undefined,
): LanguageModelV1FinishReason {
  switch (finishReason) {
    case 'stop':
      return 'stop';
    case 'length':
      return 'length';
    case 'content_filter':
      return 'content-filter';
    case 'function_call':
    case 'tool_calls':
      return 'tool-calls';
    default:
      return 'unknown';
  }
}

interface OpenAICompletionLogProps {
  tokens: string[];
  token_logprobs: number[];
  top_logprobs: Array<Record<string, number>> | null;
}

export function mapOpenAICompletionLogProbs(
  logprobs: OpenAICompletionLogProps | null | undefined,
): LanguageModelV1LogProbs | undefined {
  return logprobs?.tokens.map((token, index) => ({
    token,
    logprob: logprobs.token_logprobs[index],
    topLogprobs: logprobs.top_logprobs
      ? Object.entries(logprobs.top_logprobs[index]).map(
          ([token, logprob]) => ({
            token,
            logprob,
          }),
        )
      : [],
  }));
}

export function convertToOpenAICompletionPrompt({
  prompt,
  inputFormat,
  user = 'user',
  assistant = 'assistant',
}: {
  prompt: LanguageModelV1Prompt;
  inputFormat: 'prompt' | 'messages';
  user?: string;
  assistant?: string;
}): {
  prompt: string;
  stopSequences?: string[];
} {
  // When the user supplied a prompt input, we don't transform it:
  if (
    inputFormat === 'prompt' &&
    prompt.length === 1 &&
    prompt[0].role === 'user' &&
    prompt[0].content.length === 1 &&
    prompt[0].content[0].type === 'text'
  ) {
    return { prompt: prompt[0].content[0].text };
  }

  // otherwise transform to a chat message format:
  let text = '';

  // if first message is a system message, add it to the text:
  if (prompt[0].role === 'system') {
    text += `${prompt[0].content}\n\n`;
    prompt = prompt.slice(1);
  }

  for (const { role, content } of prompt) {
    switch (role) {
      case 'system': {
        throw new InvalidPromptError({
          message: 'Unexpected system message in prompt: ${content}',
          prompt,
        });
      }

      case 'user': {
        const userMessage = content
          .map(part => {
            switch (part.type) {
              case 'text': {
                return part.text;
              }
              case 'image': {
                throw new UnsupportedFunctionalityError({
                  functionality: 'images',
                });
              }
            }
          })
          .join('');

        text += `${user}:\n${userMessage}\n\n`;
        break;
      }

      case 'assistant': {
        const assistantMessage = content
          .map(part => {
            switch (part.type) {
              case 'text': {
                return part.text;
              }
              case 'tool-call': {
                throw new UnsupportedFunctionalityError({
                  functionality: 'tool-call messages',
                });
              }
            }
          })
          .join('');

        text += `${assistant}:\n${assistantMessage}\n\n`;
        break;
      }

      case 'tool': {
        throw new UnsupportedFunctionalityError({
          functionality: 'tool messages',
        });
      }

      default: {
        const _exhaustiveCheck: never = role;
        throw new Error(`Unsupported role: ${_exhaustiveCheck}`);
      }
    }
  }

  // Assistant message prefix:
  text += `${assistant}:\n`;

  return {
    prompt: text,
    stopSequences: [`\n${user}:`],
  };
}

interface AzureOpenAICompletionConfig {
  provider: string;
  url: (options: { modelId: string; path: string }) => string;
  compatibility: 'strict' | 'compatible';
  headers: () => Record<string, string | undefined>;
  fetch?: typeof fetch;
}

export class AzureOpenAICompletionLanguageModel implements LanguageModelV1 {
  readonly specificationVersion = 'v1';
  readonly defaultObjectGenerationMode = undefined;

  readonly modelId: OpenAICompletionModelId;
  readonly settings: OpenAICompletionSettings;

  private readonly config: AzureOpenAICompletionConfig;

  constructor(
    modelId: OpenAICompletionModelId,
    settings: OpenAICompletionSettings,
    config: AzureOpenAICompletionConfig,
  ) {
    this.modelId = modelId;
    this.settings = settings;
    this.config = config;
  }

  get provider(): string {
    return this.config.provider;
  }

  private getArgs({
    mode,
    inputFormat,
    prompt,
    maxTokens,
    temperature,
    topP,
    frequencyPenalty,
    presencePenalty,
    seed,
  }: Parameters<LanguageModelV1['doGenerate']>[0]) {
    const type = mode.type;

    const { prompt: completionPrompt, stopSequences } =
      convertToOpenAICompletionPrompt({ prompt, inputFormat });

    const baseArgs = {
      // model id:
      model: this.modelId,

      // model specific settings:
      echo: this.settings.echo,
      logit_bias: this.settings.logitBias,
      logprobs:
        typeof this.settings.logprobs === 'number'
          ? this.settings.logprobs
          : typeof this.settings.logprobs === 'boolean'
          ? this.settings.logprobs
            ? 0
            : undefined
          : undefined,
      suffix: this.settings.suffix,
      user: this.settings.user,

      // standardized settings:
      max_tokens: maxTokens,
      temperature,
      top_p: topP,
      frequency_penalty: frequencyPenalty,
      presence_penalty: presencePenalty,
      seed,

      // prompt:
      prompt: completionPrompt,

      // stop sequences:
      stop: stopSequences,
    };

    switch (type) {
      case 'regular': {
        if (mode.tools?.length) {
          throw new UnsupportedFunctionalityError({
            functionality: 'tools',
          });
        }

        if (mode.toolChoice) {
          throw new UnsupportedFunctionalityError({
            functionality: 'toolChoice',
          });
        }

        return baseArgs;
      }

      case 'object-json': {
        throw new UnsupportedFunctionalityError({
          functionality: 'object-json mode',
        });
      }

      case 'object-tool': {
        throw new UnsupportedFunctionalityError({
          functionality: 'object-tool mode',
        });
      }

      case 'object-grammar': {
        throw new UnsupportedFunctionalityError({
          functionality: 'object-grammar mode',
        });
      }

      default: {
        const _exhaustiveCheck: never = type;
        throw new Error(`Unsupported type: ${_exhaustiveCheck}`);
      }
    }
  }

  async doGenerate(
    options: Parameters<LanguageModelV1['doGenerate']>[0],
  ): Promise<Awaited<ReturnType<LanguageModelV1['doGenerate']>>> {
    const args = this.getArgs(options);

    const { responseHeaders, value: response } = await postJsonToApi({
      url: this.config.url({
        path: '/completions',
        modelId: this.modelId,
      }),
      headers: this.config.headers(),
      body: args,
      failedResponseHandler: openaiFailedResponseHandler,
      successfulResponseHandler: createJsonResponseHandler(
        openAICompletionResponseSchema,
      ),
      abortSignal: options.abortSignal,
      fetch: this.config.fetch,
    });

    const { prompt: rawPrompt, ...rawSettings } = args;
    const choice = response.choices[0];

    return {
      text: choice.text,
      usage: {
        promptTokens: response.usage.prompt_tokens,
        completionTokens: response.usage.completion_tokens,
      },
      finishReason: mapOpenAIFinishReason(choice.finish_reason),
      logprobs: mapOpenAICompletionLogProbs(choice.logprobs),
      rawCall: { rawPrompt, rawSettings },
      rawResponse: { headers: responseHeaders },
      warnings: [],
    };
  }

  async doStream(
    options: Parameters<LanguageModelV1['doStream']>[0],
  ): Promise<Awaited<ReturnType<LanguageModelV1['doStream']>>> {
    const args = this.getArgs(options);

    const { responseHeaders, value: response } = await postJsonToApi({
      url: this.config.url({
        path: '/completions',
        modelId: this.modelId,
      }),
      headers: this.config.headers(),
      body: {
        ...this.getArgs(options),
        stream: true,

        // only include stream_options when in strict compatibility mode:
        stream_options:
          this.config.compatibility === 'strict'
            ? { include_usage: true }
            : undefined,
      },
      failedResponseHandler: openaiFailedResponseHandler,
      successfulResponseHandler: createEventSourceResponseHandler(
        openaiCompletionChunkSchema,
      ),
      abortSignal: options.abortSignal,
      fetch: this.config.fetch,
    });

    const { prompt: rawPrompt, ...rawSettings } = args;

    let finishReason: LanguageModelV1FinishReason = 'other';
    let usage: { promptTokens: number; completionTokens: number } = {
      promptTokens: Number.NaN,
      completionTokens: Number.NaN,
    };
    let logprobs: LanguageModelV1LogProbs;

    return {
      stream: response.pipeThrough(
        new TransformStream<
          ParseResult<z.infer<typeof openaiCompletionChunkSchema>>,
          LanguageModelV1StreamPart
        >({
          transform(chunk, controller) {
            // handle failed chunk parsing / validation:
            if (!chunk.success) {
              finishReason = 'error';
              controller.enqueue({ type: 'error', error: chunk.error });
              return;
            }

            const value = chunk.value;

            // handle error chunks:
            if ('error' in value) {
              finishReason = 'error';
              controller.enqueue({ type: 'error', error: value.error });
              return;
            }

            if (value.usage != null) {
              usage = {
                promptTokens: value.usage.prompt_tokens,
                completionTokens: value.usage.completion_tokens,
              };
            }

            const choice = value.choices[0];

            if (choice?.finish_reason != null) {
              finishReason = mapOpenAIFinishReason(choice.finish_reason);
            }

            if (choice?.text != null) {
              controller.enqueue({
                type: 'text-delta',
                textDelta: choice.text,
              });
            }

            const mappedLogprobs = mapOpenAICompletionLogProbs(
              choice?.logprobs,
            );
            if (mappedLogprobs?.length) {
              if (logprobs === undefined) logprobs = [];
              logprobs.push(...mappedLogprobs);
            }
          },

          flush(controller) {
            controller.enqueue({
              type: 'finish',
              finishReason,
              logprobs,
              usage,
            });
          },
        }),
      ),
      rawCall: { rawPrompt, rawSettings },
      rawResponse: { headers: responseHeaders },
      warnings: [],
    };
  }
}

// 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 openAICompletionResponseSchema = z.object({
  choices: z.array(
    z.object({
      text: z.string(),
      finish_reason: z.string(),
      logprobs: z
        .object({
          tokens: z.array(z.string()),
          token_logprobs: z.array(z.number()),
          top_logprobs: z.array(z.record(z.string(), z.number())).nullable(),
        })
        .nullable()
        .optional(),
    }),
  ),
  usage: z.object({
    prompt_tokens: z.number(),
    completion_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 openaiCompletionChunkSchema = z.union([
  z.object({
    choices: z.array(
      z.object({
        text: z.string(),
        finish_reason: z.string().nullish(),
        index: z.number(),
        logprobs: z
          .object({
            tokens: z.array(z.string()),
            token_logprobs: z.array(z.number()),
            top_logprobs: z.array(z.record(z.string(), z.number())).nullable(),
          })
          .nullable()
          .optional(),
      }),
    ),
    usage: z
      .object({
        prompt_tokens: z.number(),
        completion_tokens: z.number(),
      })
      .optional()
      .nullable(),
  }),
  openAIErrorDataSchema,
]);