@pwshub/aisdk 0.0.2 → 0.0.3

package/README.md

@@ -11,6 +11,17 @@ A thin, unified AI client for OpenAI, Anthropic, Google, DashScope, and DeepSeek
 - **Token usage tracking**: Detailed token counts and estimated cost per request
 - **Provider-specific options**: Pass provider-specific parameters when needed
 
+## Limitations
+
+This package is designed for **personal project usage** with a focus on simplicity:
+
+- **Text-only chat**: Supports basic text generation and conversation
+- **No streaming**: All responses are returned as complete results
+- **No multimodal inputs**: Images, audio, video, and file uploads are not supported
+- **No function calling**: Tool use and function calling features are not available
+
+For production applications requiring advanced features, consider using the official provider SDKs directly.
+
 ## Installation
 
 ```bash
@@ -80,7 +91,8 @@ Sends a text generation request.
     inputTokens: number,
     outputTokens: number,
     cacheTokens: number,
-    estimatedCost: number // USD
+    reasoningTokens: number,  // Reasoning/thinking tokens (0 for non-reasoning models)
+    estimatedCost: number     // USD
   }
 }
 ```
@@ -133,6 +145,27 @@ const result = await ai.ask({
 })
 ```
 
+### Google (Disable Thinking Mode)
+
+Gemini 2.5 Pro and other reasoning models use thinking tokens by default. Disable thinking mode to reduce latency and cost:
+
+```javascript
+const result = await ai.ask({
+  model: 'gemini-2.5-pro',
+  apikey: process.env.GOOGLE_API_KEY,
+  prompt: 'What is the capital of Vietnam?',
+  maxTokens: 256,
+  providerOptions: {
+    thinkingConfig: {
+      thinkingBudget: 0,      // Disable reasoning tokens
+      includeThoughts: false, // Don't include thought process in response
+    },
+  },
+})
+```
+
+> **Note:** When thinking mode is enabled (default for Gemini 2.5 Pro), the model may use most of the `maxTokens` budget for reasoning. Set a higher `maxTokens` (e.g., 2048) or disable thinking with `thinkingBudget: 0`.
+
 ### With Fallbacks
 
 ```javascript
@@ -166,6 +199,36 @@ const result = await ai.ask({
 })
 ```
 
+### DashScope with Custom Region
+
+DashScope endpoints vary by region. Use `gatewayUrl` to specify your region:
+
+```javascript
+import { createAi } from '@pwshub/aisdk'
+
+// Singapore region
+const aiSingapore = createAi({
+  gatewayUrl: 'https://dashscope-intl.aliyuncs.com/compatible-mode/v1',
+})
+
+// Virginia region (US)
+const aiUS = createAi({
+  gatewayUrl: 'https://dashscope-us.aliyuncs.com/compatible-mode/v1',
+})
+
+// Beijing region (China)
+const aiCN = createAi({
+  gatewayUrl: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
+})
+
+// Use the regional client
+const result = await aiSingapore.ask({
+  model: 'qwen3.5-plus',
+  apikey: process.env.DASHSCOPE_API_KEY,
+  prompt: 'Hello from Singapore!',
+})
+```
+
 ### DeepSeek
 
 ```javascript

package/index.d.ts

@@ -25,6 +25,7 @@ export interface Usage {
   inputTokens: number;
   outputTokens: number;
   cacheTokens: number;
+  reasoningTokens: number;
   estimatedCost: number;
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pwshub/aisdk",
-  "version": "0.0.2",
+  "version": "0.0.3",
   "description": "A thin, unified AI client for OpenAI, Anthropic, Google, DashScope, and DeepSeek with automatic param normalization and fallback support",
   "repository": {
     "type": "git",
@@ -11,11 +11,12 @@
     "bun": ">=1.0.0"
   },
   "type": "module",
-  "main": "./src/index.js",
   "exports": {
-    ".": "./src/index.js"
+    ".": {
+      "types": "./index.d.ts",
+      "default": "./src/index.js"
+    }
   },
-  "types": "./index.d.ts",
   "files": [
     "src",
     "index.d.ts"

package/src/index.js

@@ -12,7 +12,7 @@
  *   temperature: 0.5,
  * })
  * console.log(result.text)
- * console.log(result.usage) // { inputTokens, outputTokens, cacheTokens, estimatedCost }
+ * console.log(result.usage) // { inputTokens, outputTokens, cacheTokens, reasoningTokens, estimatedCost }
  *
  * @example With fallbacks
  * const result = await ai.ask({
@@ -81,6 +81,7 @@ export {
  * @property {number} inputTokens
  * @property {number} outputTokens
  * @property {number} cacheTokens
+ * @property {number} reasoningTokens
  * @property {number} estimatedCost   - In USD, based on models.json pricing
  */
 
@@ -113,7 +114,7 @@ const extractGenConfig = (params) => {
 const calcCost = (usage, record) => {
   const M = 1_000_000
   const inputCost = (usage.inputTokens / M) * record.input_price
-  const outputCost = (usage.outputTokens / M) * record.output_price
+  const outputCost = ((usage.outputTokens + usage.reasoningTokens) / M) * record.output_price
   const cacheCost = (usage.cacheTokens / M) * record.cache_price
 
   // Round to 8 decimal places to avoid floating point noise

package/src/providers.js

@@ -23,7 +23,8 @@
  * @typedef {Object} RawUsage
  * @property {number} inputTokens
  * @property {number} outputTokens
- * @property {number} cacheTokens   - 0 when not applicable
+ * @property {number} cacheTokens     - 0 when not applicable
+ * @property {number} reasoningTokens - 0 when not applicable
  */
 
 /**
@@ -84,6 +85,7 @@ const openai = {
     inputTokens: data.usage?.prompt_tokens ?? 0,
     outputTokens: data.usage?.completion_tokens ?? 0,
     cacheTokens: data.usage?.prompt_tokens_details?.cached_tokens ?? 0,
+    reasoningTokens: data.usage?.completion_tokens_details?.reasoning_tokens ?? 0,
   }),
 }
 
@@ -119,7 +121,8 @@ const anthropic = {
   extractUsage: (data) => ({
     inputTokens: data.usage?.input_tokens ?? 0,
     outputTokens: data.usage?.output_tokens ?? 0,
-    cacheTokens: data.usage?.cache_read_input_tokens ?? 0,
+    cacheTokens: (data.usage?.cache_read_input_tokens ?? 0) + (data.usage?.cache_creation_input_tokens ?? 0),
+    reasoningTokens: 0,
   }),
 }
 
@@ -155,17 +158,53 @@ const google = {
       throw new Error('Google response blocked by safety filters')
     }
 
-    const text = candidate.content?.parts?.[0]?.text
-    if (!text) {
+    // Handle different content structures
+    const content = candidate.content
+    if (!content) {
       throw new Error('Google response missing content')
     }
-    return text
+
+    // Gemini 2.5 Pro may return parts as array or direct text
+    if (Array.isArray(content.parts)) {
+      const text = content.parts[0]?.text
+      if (!text) {
+        // Model may have used all tokens for reasoning (thoughtsTokenCount)
+        const thoughts = data.usageMetadata?.thoughtsTokenCount ?? 0
+        if (finishReason === 'MAX_TOKENS' && thoughts > 0) {
+          throw new Error(`Google response missing content (used ${thoughts} tokens for reasoning, maxTokens may be too low)`)
+        }
+        throw new Error('Google response missing content')
+      }
+      return text
+    }
+
+    // Some models may return content directly
+    if (typeof content.parts === 'string') {
+      return content.parts
+    }
+
+    throw new Error('Google response missing content')
+  },
+  extractUsage: (data) => {
+    // For Gemini models with reasoning, candidatesTokenCount may be undefined
+    // when all tokens were used for thinking. Calculate output tokens from
+    // totalTokenCount - promptTokenCount to get actual tokens used.
+    const totalTokens = data.usageMetadata?.totalTokenCount ?? 0
+    const promptTokens = data.usageMetadata?.promptTokenCount ?? 0
+    const candidatesTokens = data.usageMetadata?.candidatesTokenCount ?? 0
+    const thoughtsTokens = data.usageMetadata?.thoughtsTokenCount ?? 0
+    
+    // outputTokens = actual generated tokens (including reasoning)
+    // If candidatesTokenCount is missing, derive from total - prompt
+    const outputTokens = candidatesTokens || (totalTokens - promptTokens)
+    
+    return {
+      inputTokens: promptTokens,
+      outputTokens,
+      cacheTokens: data.usageMetadata?.cachedContentTokenCount ?? 0,
+      reasoningTokens: thoughtsTokens,
+    }
   },
-  extractUsage: (data) => ({
-    inputTokens: data.usageMetadata?.promptTokenCount ?? 0,
-    outputTokens: data.usageMetadata?.candidatesTokenCount ?? 0,
-    cacheTokens: data.usageMetadata?.cachedContentTokenCount ?? 0,
-  }),
 }
 
 /** @type {ProviderAdapter} */
@@ -198,6 +237,7 @@ const dashscope = {
       inputTokens: usage?.input_tokens ?? usage?.prompt_tokens ?? 0,
       outputTokens: usage?.output_tokens ?? usage?.completion_tokens ?? 0,
       cacheTokens: 0,
+      reasoningTokens: 0,
     }
   },
 }
@@ -225,7 +265,8 @@ const deepseek = {
   extractUsage: (data) => ({
     inputTokens: data.usage?.prompt_tokens ?? 0,
     outputTokens: data.usage?.completion_tokens ?? 0,
-    cacheTokens: 0,
+    cacheTokens: data.usage?.prompt_cache_hit_tokens ?? 0,
+    reasoningTokens: data.usage?.completion_tokens_details?.reasoning_tokens ?? 0,
   }),
 }