feat: add reasoning token support (#4462)

Co-authored-by: Lars Grammel <lars.grammel@gmail.com>
vercel · Jan 22, 2025 · 0a699f1 · 0a699f1
1 parent ca1b5cd
commit 0a699f1
Show file tree

Hide file tree

Showing 30 changed files with 887 additions and 13 deletions.
diff --git a/.changeset/tiny-rice-obey.md b/.changeset/tiny-rice-obey.md
@@ -0,0 +1,9 @@
+---
+'@ai-sdk/openai-compatible': patch
+'@ai-sdk/deepseek': patch
+'@ai-sdk/provider': patch
+'@ai-sdk/ui-utils': patch
+'ai': patch
+---
+
+feat: add reasoning token support
diff --git a/content/docs/02-foundations/02-providers-and-models.mdx b/content/docs/02-foundations/02-providers-and-models.mdx
@@ -105,6 +105,7 @@ Here are the capabilities of popular models:
 | [xAI Grok](/providers/ai-sdk-providers/xai)                              | `grok-beta`                  | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [xAI Grok](/providers/ai-sdk-providers/xai)                              | `grok-vision-beta`           | <Check size={18} /> | <Cross size={18} /> | <Cross size={18} /> | <Cross size={18} /> |
 | [DeepSeek](/providers/ai-sdk-providers/deepseek)                         | `deepseek-chat`              | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
+| [DeepSeek](/providers/ai-sdk-providers/deepseek)                         | `deepseek-reasoner`          | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [Cerebras](/providers/ai-sdk-providers/cerebras)                         | `llama3.1-8b`                | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [Cerebras](/providers/ai-sdk-providers/cerebras)                         | `llama3.1-70b`               | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [Cerebras](/providers/ai-sdk-providers/cerebras)                         | `llama3.3-70b`               | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |

diff --git a/content/docs/03-ai-sdk-core/05-generating-text.mdx b/content/docs/03-ai-sdk-core/05-generating-text.mdx
@@ -47,6 +47,7 @@ The result object of `generateText` contains several promises that resolve when
 - `result.text`: The generated text.
 - `result.finishReason`: The reason the model finished generating text.
 - `result.usage`: The usage of the model during text generation.
+- `result.reasoning`: The reasoning text of the model (only available for some models).
 
 ## `streamText`
 
@@ -100,6 +101,7 @@ When using `streamText`, you can provide an `onChunk` callback that is triggered
 It receives the following chunk types:
 
 - `text-delta`
+- `reasoning`
 - `tool-call`
 - `tool-result`
 - `tool-call-streaming-start` (when `experimental_toolCallStreaming` is enabled)
@@ -170,6 +172,10 @@ for await (const part of result.fullStream) {
       // handle text delta here
       break;
     }
+    case 'reasoning': {
+      // handle reasoning here
+      break;
+    }
     case 'tool-call': {
       switch (part.toolName) {
         case 'cityAttractions': {

diff --git a/content/docs/04-ai-sdk-ui/50-stream-protocol.mdx b/content/docs/04-ai-sdk-ui/50-stream-protocol.mdx
@@ -111,6 +111,14 @@ Example: `0:"example"\n`
   height={1148}
 />
 
+### Reasoning Part
+
+The reasoning parts are appended to the message as they are received. The reasoning part is available through `reasoning`.
+
+Format: `g:string\n`
+
+Example: `g:"I will open the conversation with witty banter."\n`
+
 ### Data Part
 
 The data parts are parsed as JSON and appended to the message as they are received. The data part is available through `data`.

diff --git a/content/docs/07-reference/01-ai-sdk-core/01-generate-text.mdx b/content/docs/07-reference/01-ai-sdk-core/01-generate-text.mdx
@@ -698,6 +698,12 @@ To see `generateText` in action, check out [these examples](#examples).
       type: 'string',
       description: 'The generated text by the model.',
     },
+    {
+      name: 'reasoning',
+      type: 'string | undefined',
+      description:
+        'The reasoning text of the model (only available for some models).',
+    },
     {
       name: 'toolCalls',
       type: 'array',
@@ -841,6 +847,12 @@ To see `generateText` in action, check out [these examples](#examples).
               type: 'string',
               description: 'The generated text by the model.',
             },
+            {
+              name: 'reasoning',
+              type: 'string | undefined',
+              description:
+                'The reasoning text of the model (only available for some models).',
+            },
             {
               name: 'toolCalls',
               type: 'array',

diff --git a/content/docs/07-reference/01-ai-sdk-core/02-stream-text.mdx b/content/docs/07-reference/01-ai-sdk-core/02-stream-text.mdx
@@ -603,6 +603,22 @@ To see `streamText` in action, check out [these examples](#examples).
                     },
                   ],
                 },
+                {
+                  type: 'TextStreamPart',
+                  parameters: [
+                    {
+                      name: 'type',
+                      type: "'reasoning'",
+                      description:
+                        'The type to identify the object as reasoning.',
+                    },
+                    {
+                      name: 'textDelta',
+                      type: 'string',
+                      description: 'The reasoning text delta.',
+                    },
+                  ],
+                },
                 {
                   type: 'TextStreamPart',
                   parameters: [
@@ -1056,6 +1072,12 @@ To see `streamText` in action, check out [these examples](#examples).
       description:
         'The full text that has been generated. Resolved when the response is finished.',
     },
+    {
+      name: 'reasoning',
+      type: 'Promise<string | undefined>',
+      description:
+        'The reasoning text of the model (only available for some models). Resolved when the response is finished.',
+    },
     {
       name: 'toolCalls',
       type: 'Promise<ToolCall[]>',
@@ -1156,6 +1178,12 @@ To see `streamText` in action, check out [these examples](#examples).
               type: 'string',
               description: 'The generated text by the model.',
             },
+            {
+              name: 'reasoning',
+              type: 'string | undefined',
+              description:
+                'The reasoning text of the model (only available for some models).',
+            },
             {
               name: 'toolCalls',
               type: 'array',
@@ -1312,6 +1340,21 @@ To see `streamText` in action, check out [these examples](#examples).
             },
           ],
         },
+        {
+          type: 'TextStreamPart',
+          parameters: [
+            {
+              name: 'type',
+              type: "'reasoning'",
+              description: 'The type to identify the object as reasoning.',
+            },
+            {
+              name: 'textDelta',
+              type: 'string',
+              description: 'The reasoning text delta.',
+            },
+          ],
+        },
         {
           type: 'TextStreamPart',
           parameters: [

diff --git a/content/providers/01-ai-sdk-providers/30-deepseek.mdx b/content/providers/01-ai-sdk-providers/30-deepseek.mdx
@@ -81,9 +81,10 @@ DeepSeek language models can be used in the `streamText` and `streamUI` function
 
 ## Model Capabilities
 
-| Model           | Image Input         | Object Generation   | Tool Usage          | Tool Streaming      |
-| --------------- | ------------------- | ------------------- | ------------------- | ------------------- |
-| `deepseek-chat` | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
+| Model               | Image Input         | Object Generation   | Tool Usage          | Tool Streaming      |
+| ------------------- | ------------------- | ------------------- | ------------------- | ------------------- |
+| `deepseek-chat`     | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
+| `deepseek-reasoner` | <Cross size={18} /> | <Cross size={18} /> | <Cross size={18} /> | <Cross size={18} /> |
 
 <Note>
   Please see the [DeepSeek docs](https://api-docs.deepseek.com) for a full list

diff --git a/content/providers/01-ai-sdk-providers/index.mdx b/content/providers/01-ai-sdk-providers/index.mdx
@@ -44,6 +44,7 @@ Not all providers support all AI SDK features. Here's a quick comparison of the
 | [xAI Grok](/providers/ai-sdk-providers/xai)                              | `grok-beta`                  | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [xAI Grok](/providers/ai-sdk-providers/xai)                              | `grok-vision-beta`           | <Check size={18} /> | <Cross size={18} /> | <Cross size={18} /> | <Cross size={18} /> |
 | [DeepSeek](/providers/ai-sdk-providers/deepseek)                         | `deepseek-chat`              | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
+| [DeepSeek](/providers/ai-sdk-providers/deepseek)                         | `deepseek-reasoner`          | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [Cerebras](/providers/ai-sdk-providers/cerebras)                         | `llama3.1-8b`                | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [Cerebras](/providers/ai-sdk-providers/cerebras)                         | `llama3.3-70b`               | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |
 | [Groq](/providers/ai-sdk-providers/groq)                                 | `llama-3.3-70b-versatile`    | <Cross size={18} /> | <Check size={18} /> | <Check size={18} /> | <Check size={18} /> |

diff --git a/examples/ai-core/src/generate-text/deepseek.ts b/examples/ai-core/src/generate-text/deepseek.ts
@@ -4,12 +4,18 @@ import 'dotenv/config';
 
 async function main() {
   const result = await generateText({
-    model: deepseek('deepseek-chat'),
+    model: deepseek('deepseek-reasoner'),
     prompt: 'Invent a new holiday and describe its traditions.',
   });
 
+  console.log('Reasoning:');
+  console.log(result.reasoning);
+  console.log();
+
+  console.log('Text:');
   console.log(result.text);
   console.log();
+
   console.log('Token usage:', result.usage);
   console.log('Finish reason:', result.finishReason);
 }

diff --git a/examples/ai-core/src/stream-text/deepseek-fullstream.ts b/examples/ai-core/src/stream-text/deepseek-fullstream.ts
@@ -0,0 +1,38 @@
+import { deepseek } from '@ai-sdk/deepseek';
+import { streamText } from 'ai';
+import 'dotenv/config';
+
+async function main() {
+  const result = streamText({
+    model: deepseek('deepseek-reasoner'),
+    prompt: 'Invent a new holiday and describe its traditions.',
+  });
+
+  let enteredReasoning = false;
+  let enteredText = false;
+  for await (const part of result.fullStream) {
+    if (part.type === 'reasoning') {
+      if (!enteredReasoning) {
+        enteredReasoning = true;
+        console.log('\nSTREAMING REASONING:\n');
+      }
+      process.stdout.write(part.textDelta);
+    } else if (part.type === 'text-delta') {
+      if (!enteredText) {
+        enteredText = true;
+        console.log('\nSTREAMING TEXT:\n');
+      }
+      process.stdout.write(part.textDelta);
+    }
+  }
+
+  console.log();
+  console.log('\nFINAL REASONING:\n', await result.reasoning);
+  console.log('\nFINAL TEXT:\n', await result.text);
+
+  console.log();
+  console.log('Token usage:', await result.usage);
+  console.log('Finish reason:', await result.finishReason);
+}
+
+main().catch(console.error);
diff --git a/examples/ai-core/src/stream-text/deepseek-reasoning-on-chunk.ts b/examples/ai-core/src/stream-text/deepseek-reasoning-on-chunk.ts
@@ -0,0 +1,23 @@
+import { deepseek } from '@ai-sdk/deepseek';
+import { streamText } from 'ai';
+import 'dotenv/config';
+
+async function main() {
+  const result = streamText({
+    model: deepseek('deepseek-reasoner'),
+    prompt: 'Invent a new holiday and describe its traditions.',
+    onChunk({ chunk }) {
+      if (chunk.type === 'reasoning') {
+        console.log('reasoning', chunk.textDelta);
+      }
+    },
+  });
+
+  // consume stream:
+  for await (const part of result.fullStream) {
+  }
+
+  console.log('reasoning', await result.reasoning);
+}
+
+main().catch(console.error);
diff --git a/examples/ai-core/src/stream-text/deepseek.ts b/examples/ai-core/src/stream-text/deepseek.ts
@@ -4,7 +4,7 @@ import 'dotenv/config';
 
 async function main() {
   const result = streamText({
-    model: deepseek('deepseek-chat'),
+    model: deepseek('deepseek-reasoner'),
     prompt: 'Invent a new holiday and describe its traditions.',
   });