hts/packages/isdk/anthropic/anthropic-messages-language-model.ts

import {
  LanguageModelV1,
  LanguageModelV1CallWarning,
  LanguageModelV1FinishReason,
  LanguageModelV1FunctionToolCall,
  LanguageModelV1StreamPart,
  UnsupportedFunctionalityError,
} from '@ai-sdk/provider';
import { z } from 'zod';
import {
  ParseResult,
  createEventSourceResponseHandler,
  createJsonResponseHandler,
  postJsonToApi,
} from '../spec';
import { anthropicFailedResponseHandler } from './anthropic-error';
import {
  AnthropicMessagesModelId,
  AnthropicMessagesSettings,
} from './anthropic-messages-settings';
import { convertToAnthropicMessagesPrompt } from './convert-to-anthropic-messages-prompt';
import { mapAnthropicStopReason } from './map-anthropic-stop-reason';

type AnthropicMessagesConfig = {
  provider: string;
  baseUrl: string;
  headers: () => Record<string, string | undefined>;
};

export class AnthropicMessagesLanguageModel implements LanguageModelV1 {
  readonly specificationVersion = 'v1';
  readonly defaultObjectGenerationMode = 'tool';

  readonly modelId: AnthropicMessagesModelId;
  readonly settings: AnthropicMessagesSettings;

  private readonly config: AnthropicMessagesConfig;

  constructor(
    modelId: AnthropicMessagesModelId,
    settings: AnthropicMessagesSettings,
    config: AnthropicMessagesConfig,
  ) {
    this.modelId = modelId;
    this.settings = settings;
    this.config = config;
  }

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

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

    const warnings: LanguageModelV1CallWarning[] = [];

    if (frequencyPenalty != null) {
      warnings.push({
        type: 'unsupported-setting',
        setting: 'frequencyPenalty',
      });
    }

    if (presencePenalty != null) {
      warnings.push({
        type: 'unsupported-setting',
        setting: 'presencePenalty',
      });
    }

    if (seed != null) {
      warnings.push({
        type: 'unsupported-setting',
        setting: 'seed',
      });
    }

    const messagesPrompt = convertToAnthropicMessagesPrompt(prompt);

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

      // model specific settings:
      top_k: this.settings.topK,

      // standardized settings:
      max_tokens: maxTokens ?? 4096, // 4096: max model output tokens
      temperature, // uses 0..1 scale
      top_p: topP,

      // prompt:
      system: messagesPrompt.system,
      messages: messagesPrompt.messages,
    };

    switch (type) {
      case 'regular': {
        // when the tools array is empty, change it to undefined to prevent OpenAI errors:
        const tools = mode.tools?.length ? mode.tools : undefined;

        return {
          args: {
            ...baseArgs,
            tools: tools?.map(tool => ({
              name: tool.name,
              description: tool.description,
              input_schema: tool.parameters,
            })),
          },
          warnings,
        };
      }

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

      case 'object-tool': {
        const { name, description, parameters } = mode.tool;

        // add instruction to use tool:
        baseArgs.messages[baseArgs.messages.length - 1].content.push({
          type: 'text',
          text: `\n\nUse the '${name}' tool.`,
        });

        return {
          args: {
            ...baseArgs,
            tools: [{ name, description, input_schema: parameters }],
          },
          warnings,
        };
      }

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

      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, warnings } = this.getArgs(options);

    const response = await postJsonToApi({
      url: `${this.config.baseUrl}/messages`,
      headers: this.config.headers(),
      body: args,
      failedResponseHandler: anthropicFailedResponseHandler,
      successfulResponseHandler: createJsonResponseHandler(
        anthropicMessagesResponseSchema,
      ),
      abortSignal: options.abortSignal,
    });

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

    // extract text
    let text = '';
    for (const content of response.content) {
      if (content.type === 'text') {
        text += content.text;
      }
    }

    // extract tool calls
    let toolCalls: LanguageModelV1FunctionToolCall[] | undefined = undefined;
    if (response.content.some(content => content.type === 'tool_use')) {
      toolCalls = [];
      for (const content of response.content) {
        if (content.type === 'tool_use') {
          toolCalls.push({
            toolCallType: 'function',
            toolCallId: content.id,
            toolName: content.name,
            args: JSON.stringify(content.input),
          });
        }
      }
    }

    return {
      text,
      toolCalls,
      finishReason: mapAnthropicStopReason(response.stop_reason),
      usage: {
        promptTokens: response.usage.input_tokens,
        completionTokens: response.usage.output_tokens,
      },
      rawCall: { rawPrompt, rawSettings },
      warnings,
    };
  }

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

    const response = await postJsonToApi({
      url: `${this.config.baseUrl}/messages`,
      headers: this.config.headers(),
      body: {
        ...args,
        stream: true,
      },
      failedResponseHandler: anthropicFailedResponseHandler,
      successfulResponseHandler: createEventSourceResponseHandler(
        anthropicMessagesChunkSchema,
      ),
      abortSignal: options.abortSignal,
    });

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

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

    return {
      stream: response.pipeThrough(
        new TransformStream<
          ParseResult<z.infer<typeof anthropicMessagesChunkSchema>>,
          LanguageModelV1StreamPart
        >({
          transform(chunk, controller) {
            if (!chunk.success) {
              controller.enqueue({ type: 'error', error: chunk.error });
              return;
            }

            const value = chunk.value;

            switch (value.type) {
              case 'ping':
              case 'content_block_start':
              case 'content_block_stop': {
                return; // ignored
              }

              case 'content_block_delta': {
                controller.enqueue({
                  type: 'text-delta',
                  textDelta: value.delta.text,
                });
                return;
              }

              case 'message_start': {
                usage.promptTokens = value.message.usage.input_tokens;
                usage.completionTokens = value.message.usage.output_tokens;
                return;
              }

              case 'message_delta': {
                usage.completionTokens = value.usage.output_tokens;
                finishReason = mapAnthropicStopReason(value.delta.stop_reason);
                return;
              }

              case 'message_stop': {
                controller.enqueue({ type: 'finish', finishReason, usage });
                return;
              }

              default: {
                const _exhaustiveCheck: never = value;
                throw new Error(`Unsupported chunk type: ${_exhaustiveCheck}`);
              }
            }
          },
        }),
      ),
      rawCall: { rawPrompt, rawSettings },
      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 anthropicMessagesResponseSchema = z.object({
  type: z.literal('message'),
  content: z.array(
    z.discriminatedUnion('type', [
      z.object({
        type: z.literal('text'),
        text: z.string(),
      }),
      z.object({
        type: z.literal('tool_use'),
        id: z.string(),
        name: z.string(),
        input: z.unknown(),
      }),
    ]),
  ),
  stop_reason: z.string().optional().nullable(),
  usage: z.object({
    input_tokens: z.number(),
    output_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 anthropicMessagesChunkSchema = z.discriminatedUnion('type', [
  z.object({
    type: z.literal('message_start'),
    message: z.object({
      usage: z.object({
        input_tokens: z.number(),
        output_tokens: z.number(),
      }),
    }),
  }),
  z.object({
    type: z.literal('content_block_start'),
    index: z.number(),
    content_block: z.object({
      type: z.literal('text'),
      text: z.string(),
    }),
  }),
  z.object({
    type: z.literal('content_block_delta'),
    index: z.number(),
    delta: z.object({
      type: z.literal('text_delta'),
      text: z.string(),
    }),
  }),
  z.object({
    type: z.literal('content_block_stop'),
    index: z.number(),
  }),
  z.object({
    type: z.literal('message_delta'),
    delta: z.object({ stop_reason: z.string().optional().nullable() }),
    usage: z.object({ output_tokens: z.number() }),
  }),
  z.object({
    type: z.literal('message_stop'),
  }),
  z.object({
    type: z.literal('ping'),
  }),
]);