@ai-sdk/xai 4.0.7 → 4.0.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/docs/01-xai.mdx CHANGED
@@ -212,6 +212,42 @@ const { text } = await generateText({
212
212
  });
213
213
  ```
214
214
 
215
+ You can control the resolution at which the model processes an image with
216
+ the `imageDetail` provider option on the image part:
217
+
218
+ ```ts
219
+ import { xai } from '@ai-sdk/xai';
220
+ import { generateText } from 'ai';
221
+
222
+ const { text } = await generateText({
223
+ model: xai('grok-4.3'),
224
+ messages: [
225
+ {
226
+ role: 'user',
227
+ content: [
228
+ { type: 'text', text: 'What do you see in this image?' },
229
+ {
230
+ type: 'file',
231
+ mediaType: 'image/png',
232
+ data: fs.readFileSync('./image.png'),
233
+ providerOptions: {
234
+ xai: { imageDetail: 'low' },
235
+ },
236
+ },
237
+ ],
238
+ },
239
+ ],
240
+ });
241
+ ```
242
+
243
+ The following image detail values are supported:
244
+
245
+ - **low**: processes the image at reduced resolution and consumes fewer input tokens.
246
+ - **high**: processes the image at full resolution.
247
+ - **auto**: lets the xAI API decide.
248
+
249
+ When not set, the image is processed at full resolution.
250
+
215
251
  ### Web Search Tool
216
252
 
217
253
  The web search tool enables autonomous web research with optional domain filtering and image understanding:
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@ai-sdk/xai",
3
- "version": "4.0.7",
3
+ "version": "4.0.9",
4
4
  "type": "module",
5
5
  "license": "Apache-2.0",
6
6
  "sideEffects": false,
@@ -29,17 +29,17 @@
29
29
  }
30
30
  },
31
31
  "dependencies": {
32
- "@ai-sdk/openai-compatible": "3.0.5",
32
+ "@ai-sdk/openai-compatible": "3.0.6",
33
33
  "@ai-sdk/provider": "4.0.2",
34
- "@ai-sdk/provider-utils": "5.0.5"
34
+ "@ai-sdk/provider-utils": "5.0.6"
35
35
  },
36
36
  "devDependencies": {
37
37
  "@types/node": "22.19.19",
38
38
  "tsup": "^8.5.1",
39
39
  "typescript": "5.8.3",
40
40
  "zod": "3.25.76",
41
- "@ai-sdk/test-server": "2.0.0",
42
- "@vercel/ai-tsconfig": "0.0.0"
41
+ "@vercel/ai-tsconfig": "0.0.0",
42
+ "@ai-sdk/test-server": "2.0.0"
43
43
  },
44
44
  "peerDependencies": {
45
45
  "zod": "^3.25.76 || ^4.1.8"
@@ -6,15 +6,19 @@ import {
6
6
  import {
7
7
  convertToBase64,
8
8
  getTopLevelMediaType,
9
+ parseProviderOptions,
9
10
  resolveFullMediaType,
10
11
  resolveProviderReference,
11
12
  } from '@ai-sdk/provider-utils';
12
- import type { XaiChatPrompt } from './xai-chat-prompt';
13
+ import type { XaiChatPrompt, XaiUserMessageContent } from './xai-chat-prompt';
14
+ import { xaiFilePartProviderOptions } from './xai-file-part-options';
13
15
 
14
- export function convertToXaiChatMessages(prompt: LanguageModelV4Prompt): {
16
+ export async function convertToXaiChatMessages(
17
+ prompt: LanguageModelV4Prompt,
18
+ ): Promise<{
15
19
  messages: XaiChatPrompt;
16
20
  warnings: Array<SharedV4Warning>;
17
- } {
21
+ }> {
18
22
  const messages: XaiChatPrompt = [];
19
23
  const warnings: Array<SharedV4Warning> = [];
20
24
 
@@ -31,54 +35,68 @@ export function convertToXaiChatMessages(prompt: LanguageModelV4Prompt): {
31
35
  break;
32
36
  }
33
37
 
34
- messages.push({
35
- role: 'user',
36
- content: content.map(part => {
37
- switch (part.type) {
38
- case 'text': {
39
- return { type: 'text', text: part.text };
40
- }
41
- case 'file': {
42
- switch (part.data.type) {
43
- case 'reference': {
44
- return {
45
- type: 'file',
46
- file: {
47
- file_id: resolveProviderReference({
48
- reference: part.data.reference,
49
- provider: 'xai',
38
+ const userContent: Array<XaiUserMessageContent> = [];
39
+
40
+ for (const part of content) {
41
+ switch (part.type) {
42
+ case 'text': {
43
+ userContent.push({ type: 'text', text: part.text });
44
+ break;
45
+ }
46
+ case 'file': {
47
+ switch (part.data.type) {
48
+ case 'reference': {
49
+ userContent.push({
50
+ type: 'file',
51
+ file: {
52
+ file_id: resolveProviderReference({
53
+ reference: part.data.reference,
54
+ provider: 'xai',
55
+ }),
56
+ },
57
+ });
58
+ break;
59
+ }
60
+ case 'text': {
61
+ throw new UnsupportedFunctionalityError({
62
+ functionality: 'text file parts',
63
+ });
64
+ }
65
+ case 'url':
66
+ case 'data': {
67
+ if (getTopLevelMediaType(part.mediaType) === 'image') {
68
+ const filePartOptions = await parseProviderOptions({
69
+ provider: 'xai',
70
+ providerOptions: part.providerOptions,
71
+ schema: xaiFilePartProviderOptions,
72
+ });
73
+
74
+ userContent.push({
75
+ type: 'image_url',
76
+ image_url: {
77
+ url:
78
+ part.data.type === 'url'
79
+ ? part.data.url.toString()
80
+ : `data:${resolveFullMediaType({ part })};base64,${convertToBase64(part.data.data)}`,
81
+ ...(filePartOptions?.imageDetail != null && {
82
+ detail: filePartOptions.imageDetail,
50
83
  }),
51
84
  },
52
- };
53
- }
54
- case 'text': {
85
+ });
86
+ } else {
55
87
  throw new UnsupportedFunctionalityError({
56
- functionality: 'text file parts',
88
+ functionality: `file part media type ${part.mediaType}`,
57
89
  });
58
90
  }
59
- case 'url':
60
- case 'data': {
61
- if (getTopLevelMediaType(part.mediaType) === 'image') {
62
- return {
63
- type: 'image_url',
64
- image_url: {
65
- url:
66
- part.data.type === 'url'
67
- ? part.data.url.toString()
68
- : `data:${resolveFullMediaType({ part })};base64,${convertToBase64(part.data.data)}`,
69
- },
70
- };
71
- } else {
72
- throw new UnsupportedFunctionalityError({
73
- functionality: `file part media type ${part.mediaType}`,
74
- });
75
- }
76
- }
91
+ break;
77
92
  }
78
93
  }
94
+ break;
79
95
  }
80
- }),
81
- });
96
+ }
97
+ }
98
+
99
+ messages.push({ role: 'user', content: userContent });
82
100
 
83
101
  break;
84
102
  }
package/src/index.ts CHANGED
@@ -4,6 +4,7 @@ export type {
4
4
  XaiLanguageModelChatOptions as XaiProviderOptions,
5
5
  } from './xai-chat-language-model-options';
6
6
  export type { XaiErrorData } from './xai-error';
7
+ export type { XaiFilePartProviderOptions } from './xai-file-part-options';
7
8
  export type {
8
9
  XaiLanguageModelResponsesOptions,
9
10
  /** @deprecated Use `XaiLanguageModelResponsesOptions` instead. */
@@ -6,9 +6,11 @@ import {
6
6
  import {
7
7
  convertToBase64,
8
8
  getTopLevelMediaType,
9
+ parseProviderOptions,
9
10
  resolveFullMediaType,
10
11
  resolveProviderReference,
11
12
  } from '@ai-sdk/provider-utils';
13
+ import { xaiFilePartProviderOptions } from '../xai-file-part-options';
12
14
  import type {
13
15
  XaiResponsesInput,
14
16
  XaiResponsesUserMessageContentPart,
@@ -71,9 +73,18 @@ export async function convertToXaiResponsesInput({
71
73
  ? block.data.url.toString()
72
74
  : `data:${resolveFullMediaType({ part: block })};base64,${convertToBase64(block.data.data)}`;
73
75
 
76
+ const filePartOptions = await parseProviderOptions({
77
+ provider: 'xai',
78
+ providerOptions: block.providerOptions,
79
+ schema: xaiFilePartProviderOptions,
80
+ });
81
+
74
82
  contentParts.push({
75
83
  type: 'input_image',
76
84
  image_url: imageUrl,
85
+ ...(filePartOptions?.imageDetail != null && {
86
+ detail: filePartOptions.imageDetail,
87
+ }),
77
88
  });
78
89
  } else if (block.data.type === 'url') {
79
90
  // xAI's Responses API accepts non-image documents (PDF, text, CSV, etc.)
@@ -26,7 +26,11 @@ export type XaiResponsesSystemMessage = {
26
26
 
27
27
  export type XaiResponsesUserMessageContentPart =
28
28
  | { type: 'input_text'; text: string }
29
- | { type: 'input_image'; image_url: string }
29
+ | {
30
+ type: 'input_image';
31
+ image_url: string;
32
+ detail?: 'low' | 'high' | 'auto';
33
+ }
30
34
  | { type: 'input_file'; file_id: string }
31
35
  | { type: 'input_file'; file_url: string };
32
36
 
@@ -25,6 +25,7 @@ import {
25
25
  } from '@ai-sdk/provider-utils';
26
26
  import type { z } from 'zod/v4';
27
27
  import { getResponseMetadata } from '../get-response-metadata';
28
+ import { supportsReasoningEffort } from '../supports-reasoning-effort';
28
29
  import { xaiFailedResponseHandler } from '../xai-error';
29
30
  import { convertToXaiResponsesInput } from './convert-to-xai-responses-input';
30
31
  import { convertXaiResponsesUsage } from './convert-xai-responses-usage';
@@ -164,23 +165,30 @@ export class XaiResponsesLanguageModel implements LanguageModelV4 {
164
165
  }
165
166
  }
166
167
 
167
- const resolvedReasoningEffort =
168
- options.reasoningEffort ??
169
- (isCustomReasoning(reasoning)
170
- ? reasoning === 'none'
171
- ? undefined
172
- : mapReasoningToProviderEffort({
173
- reasoning,
174
- effortMap: {
175
- minimal: 'low',
176
- low: 'low',
177
- medium: 'medium',
178
- high: 'high',
179
- xhigh: 'high',
180
- },
181
- warnings,
182
- })
183
- : undefined);
168
+ let resolvedReasoningEffort = options.reasoningEffort;
169
+ if (resolvedReasoningEffort == null && isCustomReasoning(reasoning)) {
170
+ if (!supportsReasoningEffort(this.modelId)) {
171
+ warnings.push({
172
+ type: 'unsupported',
173
+ feature: 'reasoning',
174
+ details: `reasoning "${reasoning}" is not supported by this model.`,
175
+ });
176
+ } else if (reasoning === 'none') {
177
+ resolvedReasoningEffort = 'none';
178
+ } else {
179
+ resolvedReasoningEffort = mapReasoningToProviderEffort({
180
+ reasoning,
181
+ effortMap: {
182
+ minimal: 'low',
183
+ low: 'low',
184
+ medium: 'medium',
185
+ high: 'high',
186
+ xhigh: 'high',
187
+ },
188
+ warnings,
189
+ });
190
+ }
191
+ }
184
192
 
185
193
  const baseArgs: Record<string, unknown> = {
186
194
  model: this.modelId,
@@ -0,0 +1,12 @@
1
+ /**
2
+ * The grok-4.20 reasoning and non-reasoning models (including dated variants
3
+ * such as `grok-4.20-0309-reasoning`) reject the reasoning effort parameter
4
+ * with an invalid-argument error for every value, including `none`.
5
+ * Other models such as `grok-4.3`, `grok-latest`, and
6
+ * `grok-4.20-multi-agent` accept it.
7
+ */
8
+ const modelsWithoutReasoningEffort = /^grok-4\.20(-\d{4})?-(non-)?reasoning$/;
9
+
10
+ export function supportsReasoningEffort(modelId: string): boolean {
11
+ return !modelsWithoutReasoningEffort.test(modelId);
12
+ }
@@ -31,6 +31,7 @@ import { convertToXaiChatMessages } from './convert-to-xai-chat-messages';
31
31
  import { convertXaiChatUsage } from './convert-xai-chat-usage';
32
32
  import { getResponseMetadata } from './get-response-metadata';
33
33
  import { mapXaiFinishReason } from './map-xai-finish-reason';
34
+ import { supportsReasoningEffort } from './supports-reasoning-effort';
34
35
  import {
35
36
  xaiLanguageModelChatOptions,
36
37
  type XaiChatModelId,
@@ -125,7 +126,7 @@ export class XaiChatLanguageModel implements LanguageModelV4 {
125
126
 
126
127
  // convert ai sdk messages to xai format
127
128
  const { messages, warnings: messageWarnings } =
128
- convertToXaiChatMessages(prompt);
129
+ await convertToXaiChatMessages(prompt);
129
130
  warnings.push(...messageWarnings);
130
131
 
131
132
  // prepare tools for xai
@@ -139,6 +140,31 @@ export class XaiChatLanguageModel implements LanguageModelV4 {
139
140
  });
140
141
  warnings.push(...toolWarnings);
141
142
 
143
+ let reasoningEffort = options.reasoningEffort;
144
+ if (reasoningEffort == null && isCustomReasoning(reasoning)) {
145
+ if (!supportsReasoningEffort(this.modelId)) {
146
+ warnings.push({
147
+ type: 'unsupported',
148
+ feature: 'reasoning',
149
+ details: `reasoning "${reasoning}" is not supported by this model.`,
150
+ });
151
+ } else if (reasoning === 'none') {
152
+ reasoningEffort = 'none';
153
+ } else {
154
+ reasoningEffort = mapReasoningToProviderEffort({
155
+ reasoning,
156
+ effortMap: {
157
+ minimal: 'low',
158
+ low: 'low',
159
+ medium: 'medium',
160
+ high: 'high',
161
+ xhigh: 'high',
162
+ },
163
+ warnings,
164
+ });
165
+ }
166
+ }
167
+
142
168
  const baseArgs = {
143
169
  // model id
144
170
  model: this.modelId,
@@ -153,23 +179,7 @@ export class XaiChatLanguageModel implements LanguageModelV4 {
153
179
  temperature,
154
180
  top_p: topP,
155
181
  seed,
156
- reasoning_effort:
157
- options.reasoningEffort ??
158
- (isCustomReasoning(reasoning)
159
- ? reasoning === 'none'
160
- ? undefined
161
- : mapReasoningToProviderEffort({
162
- reasoning,
163
- effortMap: {
164
- minimal: 'low',
165
- low: 'low',
166
- medium: 'medium',
167
- high: 'high',
168
- xhigh: 'high',
169
- },
170
- warnings,
171
- })
172
- : undefined),
182
+ reasoning_effort: reasoningEffort,
173
183
 
174
184
  // parallel function calling
175
185
  parallel_function_calling: options.parallel_function_calling,
@@ -18,7 +18,10 @@ export interface XaiUserMessage {
18
18
 
19
19
  export type XaiUserMessageContent =
20
20
  | { type: 'text'; text: string }
21
- | { type: 'image_url'; image_url: { url: string } }
21
+ | {
22
+ type: 'image_url';
23
+ image_url: { url: string; detail?: 'low' | 'high' | 'auto' };
24
+ }
22
25
  | { type: 'file'; file: { file_id: string } };
23
26
 
24
27
  export interface XaiAssistantMessage {
@@ -0,0 +1,21 @@
1
+ import { z } from 'zod/v4';
2
+
3
+ // provider options for file parts (images) in user messages
4
+ export const xaiFilePartProviderOptions = z.object({
5
+ /**
6
+ * Controls the resolution at which the model processes the image.
7
+ * `low` processes the image at reduced resolution and consumes fewer
8
+ * input tokens, `high` processes the image at full resolution, and
9
+ * `auto` lets the API decide. Defaults to full resolution when not set.
10
+ *
11
+ * Note: the xAI API silently ignores invalid values, so the value is
12
+ * validated client-side.
13
+ *
14
+ * @see https://docs.x.ai/developers/model-capabilities/images/understanding
15
+ */
16
+ imageDetail: z.enum(['low', 'high', 'auto']).optional(),
17
+ });
18
+
19
+ export type XaiFilePartProviderOptions = z.infer<
20
+ typeof xaiFilePartProviderOptions
21
+ >;