fix: expose deepseek cache usage #4333

packages/deepseek/src/deepseek-provider.ts

@@ -3,6 +3,7 @@ import {
   LanguageModelV1,
   NoSuchModelError,
   ProviderV1,
+  LanguageModelV1ProviderMetadata,
 } from '@ai-sdk/provider';
 import {
   FetchFunction,
@@ -13,6 +14,7 @@ import {
   DeepSeekChatModelId,
   DeepSeekChatSettings,
 } from './deepseek-chat-settings';
+import { z } from 'zod';
 
 export interface DeepSeekProviderSettings {
   /**
@@ -85,6 +87,29 @@ export function createDeepSeek(
       headers: getHeaders,
       fetch: options.fetch,
       defaultObjectGenerationMode: 'json',
+      usageStructure: z
+        .object({
+          prompt_tokens: z.number().nullish(),
+          completion_tokens: z.number().nullish(),
+          prompt_cache_hit_tokens: z.number().nullish(),
+          prompt_cache_miss_tokens: z.number().nullish(),
+        })
+        .nullish(),
+      getProviderMetadata(
+        value: any,
+        _cur: LanguageModelV1ProviderMetadata | undefined,
+      ) {
+        if (value?.usage?.prompt_cache_hit_tokens != null) {
+          return {
+            deepseek: {
+              promptCacheHitTokens: value.usage.prompt_cache_hit_tokens,
+              promptCacheMissTokens: value.usage.prompt_cache_miss_tokens,
+            },
+          };
+        } else {
+          return undefined;
+        }
+      },
     });
   };
 

packages/openai-compatible/src/openai-compatible-chat-language-model.ts

@@ -41,6 +41,13 @@ export type OpenAICompatibleChatConfig = {
   fetch?: FetchFunction;
   errorStructure?: ProviderErrorStructure<any>;
 
+  usageStructure?: z.ZodType;
+
+  getProviderMetadata?: (
+    value: any,
+    cur: LanguageModelV1ProviderMetadata | undefined,
+  ) => LanguageModelV1ProviderMetadata | undefined;
+
   /**
 Default object generation mode that should be used with this model when
 no mode is specified. Should be the mode with the best results for this
@@ -80,6 +87,7 @@ export class OpenAICompatibleChatLanguageModel implements LanguageModelV1 {
       config.errorStructure ?? defaultOpenAICompatibleErrorStructure;
     this.chunkSchema = createOpenAICompatibleChatChunkSchema(
       errorStructure.errorSchema,
+      config.usageStructure,
     );
     this.failedResponseHandler = createJsonErrorResponseHandler(errorStructure);
 
@@ -360,6 +368,7 @@ export class OpenAICompatibleChatLanguageModel implements LanguageModelV1 {
     let isFirstChunk = true;
 
     let providerMetadata: LanguageModelV1ProviderMetadata | undefined;
+    const getProviderMetadata = this.config.getProviderMetadata;
     return {
       stream: response.pipeThrough(
         new TransformStream<
@@ -392,6 +401,10 @@ export class OpenAICompatibleChatLanguageModel implements LanguageModelV1 {
               });
             }
 
+            if (typeof getProviderMetadata === 'function') {
+              providerMetadata = getProviderMetadata(value, providerMetadata);
+            }
+
             if (value.usage != null) {
               usage = {
                 promptTokens: value.usage.prompt_tokens ?? undefined,
@@ -590,6 +603,7 @@ const OpenAICompatibleChatResponseSchema = z.object({
 // this approach limits breakages when the API changes and increases efficiency
 const createOpenAICompatibleChatChunkSchema = <ERROR_SCHEMA extends z.ZodType>(
   errorSchema: ERROR_SCHEMA,
+  usageSchema?: z.ZodType,
 ) =>
   z.union([
     z.object({
@@ -620,12 +634,14 @@ const createOpenAICompatibleChatChunkSchema = <ERROR_SCHEMA extends z.ZodType>(
           finish_reason: z.string().nullish(),
         }),
       ),
-      usage: z
-        .object({
-          prompt_tokens: z.number().nullish(),
-          completion_tokens: z.number().nullish(),
-        })
-        .nullish(),
+      usage:
+        usageSchema ??
+        z
+          .object({
+            prompt_tokens: z.number().nullish(),
+            completion_tokens: z.number().nullish(),
+          })
+          .nullish(),
     }),
     errorSchema,
   ]);