feat(ai-mistral): support document (PDF) input via document_url

.changeset/mistral-document-input.md

@@ -0,0 +1,5 @@
+---
+'@tanstack/ai-mistral': minor
+---
+
+Support document (PDF) input for Mistral vision models via `document_url`. The text adapter now maps `document` content parts to Mistral's `document_url` format (hosted URLs pass through; inline bytes are wrapped in a `data:` URL, mirroring image handling), and `document` is declared as an input modality for the vision-capable models. Previously any `document` content part threw "Supported types: text, image".

packages/ai-mistral/src/adapters/text.ts

@@ -1019,8 +1019,17 @@ export class MistralTextAdapter<
       }
     }
 
+    if (part.type === 'document') {
+      const documentValue = part.source.value
+      const documentUrl =
+        part.source.type === 'data' && !documentValue.startsWith('data:')
+          ? `data:${part.source.mimeType};base64,${documentValue}`
+          : documentValue
+      return { type: 'document_url', documentUrl }
+    }
+
     throw new Error(
-      `Mistral text adapter does not support content part of type '${(part as ContentPart).type}'. Supported types: text, image. Use a vision-capable model (pixtral-large-latest, pixtral-12b-2409, mistral-medium-latest, or mistral-small-latest) for images.`,
+      `Mistral text adapter does not support content part of type '${(part as ContentPart).type}'. Supported types: text, image, document. Use a vision-capable model (mistral-medium-latest or mistral-small-latest) for images and documents.`,
     )
   }
 

packages/ai-mistral/src/model-meta.ts

@@ -19,7 +19,7 @@ interface ModelMeta<TProviderOptions = unknown> {
     output?: { normal: number }
   }
   supports: {
-    input: Array<'text' | 'image' | 'audio'>
+    input: Array<'text' | 'image' | 'audio' | 'document'>
     output: Array<'text'>
     endpoints: Array<'chat' | 'embeddings'>
 
@@ -61,7 +61,7 @@ const MISTRAL_MEDIUM_LATEST = {
     output: { normal: 2 },
   },
   supports: {
-    input: ['text', 'image'],
+    input: ['text', 'image', 'document'],
     output: ['text'],
     endpoints: ['chat'],
     features: ['streaming', 'tools', 'json_object', 'json_schema', 'vision'],
@@ -77,7 +77,7 @@ const MISTRAL_SMALL_LATEST = {
     output: { normal: 0.3 },
   },
   supports: {
-    input: ['text', 'image'],
+    input: ['text', 'image', 'document'],
     output: ['text'],
     endpoints: ['chat'],
     features: ['streaming', 'tools', 'json_object', 'json_schema', 'vision'],
@@ -141,7 +141,7 @@ const PIXTRAL_LARGE_LATEST = {
     output: { normal: 6 },
   },
   supports: {
-    input: ['text', 'image'],
+    input: ['text', 'image', 'document'],
     output: ['text'],
     endpoints: ['chat'],
     features: ['streaming', 'tools', 'json_object', 'json_schema', 'vision'],
@@ -157,7 +157,7 @@ const PIXTRAL_12B_2409 = {
     output: { normal: 0.15 },
   },
   supports: {
-    input: ['text', 'image'],
+    input: ['text', 'image', 'document'],
     output: ['text'],
     endpoints: ['chat'],
     features: ['streaming', 'tools', 'json_object', 'vision'],

packages/ai-mistral/tests/document-input.test.ts

@@ -0,0 +1,64 @@
+import { describe, expect, it } from 'vitest'
+import { createMistralText } from '../src/adapters/text'
+import type { DocumentPart } from '@tanstack/ai'
+
+// `convertContentPartToMistral` is private; reach it directly so the test
+// exercises the document mapping without standing up an HTTP mock. The
+// `messageToWire` step then snake-cases `documentUrl` -> `document_url`.
+const convertContentPart = (
+  adapter: ReturnType<typeof createMistralText>,
+  part: DocumentPart,
+): unknown =>
+  (
+    adapter as unknown as {
+      convertContentPartToMistral: (p: DocumentPart) => unknown
+    }
+  ).convertContentPartToMistral(part)
+
+describe('mistral document content parts', () => {
+  const adapter = createMistralText('mistral-small-latest', 'test-key')
+
+  it('maps a hosted-URL document to a document_url part', () => {
+    const part: DocumentPart = {
+      type: 'document',
+      source: { type: 'url', value: 'https://example.com/contract.pdf' },
+    }
+
+    expect(convertContentPart(adapter, part)).toEqual({
+      type: 'document_url',
+      documentUrl: 'https://example.com/contract.pdf',
+    })
+  })
+
+  it('wraps inline base64 document bytes in a data: URL', () => {
+    const part: DocumentPart = {
+      type: 'document',
+      source: {
+        type: 'data',
+        value: 'JVBERi0xLjQ=',
+        mimeType: 'application/pdf',
+      },
+    }
+
+    expect(convertContentPart(adapter, part)).toEqual({
+      type: 'document_url',
+      documentUrl: 'data:application/pdf;base64,JVBERi0xLjQ=',
+    })
+  })
+
+  it('passes through a data: URL that is already formatted', () => {
+    const part: DocumentPart = {
+      type: 'document',
+      source: {
+        type: 'data',
+        value: 'data:application/pdf;base64,JVBERi0xLjQ=',
+        mimeType: 'application/pdf',
+      },
+    }
+
+    expect(convertContentPart(adapter, part)).toEqual({
+      type: 'document_url',
+      documentUrl: 'data:application/pdf;base64,JVBERi0xLjQ=',
+    })
+  })
+})