feat:support image dataset search

document/content/guide/build/workflow/nodes/dataset_search.en.mdx

@@ -21,6 +21,12 @@ For detailed parameters and internal logic, see: [FastGPT Knowledge Base Search]
 
 Select one or more Knowledge Bases using the **same embedding model** for vector search.
 
+### Input - Search Content
+
+Search content is an `Array<string>`. New nodes reference the user's question by default. You can also add file links to the same input for mixed text-image retrieval.
+
+At runtime, FastGPT separates text from file links automatically: plain text goes into text search, image links go into image search, and non-image file links such as PDF, docx, xlsx, txt, and pptx are filtered out. Filtered file links are not searched as text and do not make the node fail.
+
 ### Input - Search Parameters
 
 [View parameter details](../../../dataset/dataset_engine.en.mdx#搜索参数)

document/content/guide/build/workflow/nodes/dataset_search.mdx

@@ -21,6 +21,12 @@ description: FastGPT AI 知识库搜索模块介绍
 
 可以选择一个或多个**相同向量模型**的知识库，用于向量搜索。
 
+### 输入 - 检索内容
+
+检索内容为`Array<string>`，新建节点默认引用用户问题。你也可以在同一个输入里追加文件链接，用于图文混合检索。
+
+运行时会自动拆分文本和文件链接：普通文本进入文本检索；图片链接进入图片检索；PDF、docx、xlsx、txt、pptx 等非图片文件链接会被过滤，不会作为文本参与检索，也不会让节点报错。
+
 ### 输入 - 搜索参数
 
 [点击查看参数介绍](../../../dataset/dataset_engine.mdx#搜索参数)

document/content/guide/dataset/dataset_engine.en.mdx

@@ -63,6 +63,18 @@ This means you can continuously improve data chunk accuracy through annotation.
 
 ![](/imgs/dataset_search_process.png)
 
+### Image and Mixed Text-Image Search
+
+When `Support Image Recognition` is enabled for an embedding model in model settings, FastGPT treats it as a multimodal embedding model. Image data can generate an `imageEmbedding` vector index for image-to-image search. If the Knowledge Base also has an image understanding model, images can also generate text-description indexes so text queries can match image content.
+
+Search automatically branches based on the Knowledge Base configuration:
+
+- Multimodal embedding model: text queries use text vector search, image queries use image vector search, and mixed text-image queries are merged with RRF.
+- Multimodal embedding model + image understanding model: in addition to image vector search, query images are captioned and used for text search.
+- Text-only embedding model + image understanding model: query images are captioned first, then searched as text.
+- Text-only embedding model without an image understanding model: image inputs are ignored; mixed text-image queries use only the text portion.
+
+The search test page supports local image uploads for image-only and mixed text-image tests, up to 10 images. Images uploaded for search testing are temporary, expire after 3 hours, and are not saved as Knowledge Base data.
 
 ## Search Parameters
 | | | |

document/content/guide/dataset/dataset_engine.mdx

@@ -63,6 +63,18 @@ FastGPT 采用了`PostgresSQL`的`PG Vector`插件作为向量检索器，索引
 
 ![](/imgs/dataset_search_process.png)
 
+### 图片与图文混合检索
+
+当索引模型在模型配置中开启`支持图片识别`后，FastGPT 会把它视为多模态索引模型：图片数据可以生成`imageEmbedding`图片向量索引，用于以图搜图。若知识库同时配置了图片理解模型，图片还会生成文本描述索引，便于通过文字命中图片内容。
+
+搜索时会按知识库能力自动分支：
+
+- 多模态索引模型：文本查询走文本向量检索，图片查询走图片向量检索，图文混合时通过 RRF 合并结果。
+- 多模态索引模型 + 图片理解模型：除图片向量检索外，查询图片还会先转成文本描述，再参与文本检索。
+- 普通索引模型 + 图片理解模型：图片会先转成文本描述，再走普通文本检索。
+- 普通索引模型且无图片理解模型：图片输入不会参与检索；图文混合时仅使用文字部分。
+
+搜索测试页支持本地上传图片进行纯图或图文混合测试，最多 10 张。测试上传的图片只用于临时检索，3 小时后过期，不会作为知识库正式数据保存。
 
 ## 搜索参数
 | | | |

document/content/openapi/dataset.en.mdx

@@ -1248,6 +1248,9 @@ curl --location --request POST 'http://localhost:3000/api/core/dataset/searchTes
 --data-raw '{
     "datasetId": "Dataset ID",
     "text": "Who is the director",
+    "queryImageUrls": [
+      "temp/teamId/search-image.png"
+    ],
     "limit": 5000,
     "similarity": 0,
     "searchMode": "embedding",
@@ -1265,7 +1268,8 @@ curl --location --request POST 'http://localhost:3000/api/core/dataset/searchTes
 
 <div>
 - datasetId - Dataset ID
-- text - Text to test
+- text - Text to test. It can be empty if queryImageUrls contains at least one image.
+- queryImageUrls - Temporary image keys for search testing, up to 10 images. Supports image-only search and mixed text-image search. Public image URLs, dataset keys, and chat keys are not supported here. For API calls, first upload the local image through `/api/core/dataset/file/uploadSearchTestImage`, then pass the returned `key` (`temp/${teamId}/...`) in this field. The temporary key expires after 3 hours.
 - limit - Maximum tokens
 - similarity - Minimum similarity (0~1, optional)
 - searchMode - Search mode: embedding | fullTextRecall | mixedRecall
@@ -1303,3 +1307,29 @@ Returns top k results. limit is the maximum tokens, up to 20000 tokens.
 
   </Tab>
 </Tabs>
+
+### Upload Search Test Image
+
+Use this endpoint to upload a local image from the search test page or an OpenAPI client. Only image files are supported; PDF, docx, xlsx, txt, pptx, and other document files are not accepted. Image format and size limits follow the system upload configuration. Uploaded objects are only used for temporary retrieval and expire after 3 hours.
+
+To use image retrieval with `/api/core/dataset/searchTest`, call this endpoint first and pass `data.key` from the response to `queryImageUrls`. Public image URLs are not supported directly in `queryImageUrls`.
+
+```bash
+curl --location --request POST 'http://localhost:3000/api/core/dataset/file/uploadSearchTestImage' \
+--header 'Authorization: Bearer fastgpt-xxxxx' \
+--form 'datasetId="Dataset ID"' \
+--form 'file=@"./query.png"'
+```
+
+Response example:
+
+```json
+{
+  "code": 200,
+  "statusText": "",
+  "data": {
+    "key": "temp/teamId/query.png",
+    "previewUrl": "https://..."
+  }
+}
+```

document/content/openapi/dataset.mdx

@@ -1230,6 +1230,9 @@ curl --location --request POST 'http://localhost:3000/api/core/dataset/searchTes
 --data-raw '{
     "datasetId": "知识库的ID",
     "text": "导演是谁",
+    "queryImageUrls": [
+      "temp/teamId/search-image.png"
+    ],
     "limit": 5000,
     "similarity": 0,
     "searchMode": "embedding",
@@ -1247,7 +1250,8 @@ curl --location --request POST 'http://localhost:3000/api/core/dataset/searchTes
 
 <div>
 - datasetId - 知识库ID
-- text - 需要测试的文本
+- text - 需要测试的文本，可为空；当 text 为空时，queryImageUrls 至少需要传 1 张图片
+- queryImageUrls - 搜索测试图片临时 key，最多 10 张；支持纯图片搜索和图文混合搜索。当前不支持直接传公网 URL、dataset key 或 chat key。API 调用时需要先通过 `/api/core/dataset/file/uploadSearchTestImage` 上传本地图片，使用响应中的 `key`（格式为 `temp/${teamId}/...`）填入该字段。该临时 key 3 小时后过期
 - limit - 最大 tokens 数量
 - similarity - 最低相关度（0~1，可选）
 - searchMode - 搜索模式：embedding | fullTextRecall | mixedRecall
@@ -1285,3 +1289,29 @@ curl --location --request POST 'http://localhost:3000/api/core/dataset/searchTes
 
   </Tab>
 </Tabs>
+
+### 上传搜索测试图片
+
+该接口用于搜索测试页或 OpenAPI 调用方上传本地图片。仅支持图片文件，不支持 PDF、docx、xlsx、txt、pptx 等普通文件；图片格式和大小限制跟随系统上传配置。上传对象只用于临时检索，过期时间固定为 3 小时。
+
+如果要在 `/api/core/dataset/searchTest` 中使用图片检索，先调用本接口上传图片，然后将响应中的 `data.key` 填入 `queryImageUrls`。当前不支持把公网图片 URL 直接填入 `queryImageUrls`。
+
+```bash
+curl --location --request POST 'http://localhost:3000/api/core/dataset/file/uploadSearchTestImage' \
+--header 'Authorization: Bearer fastgpt-xxxxx' \
+--form 'datasetId="知识库的ID"' \
+--form 'file=@"./query.png"'
+```
+
+响应示例：
+
+```json
+{
+  "code": 200,
+  "statusText": "",
+  "data": {
+    "key": "temp/teamId/query.png",
+    "previewUrl": "https://..."
+  }
+}
+```