Try document parsing, image content extraction, document chunking, and other services-OpenSearch(Open Search)-阿里云帮助中心

The Service Marketplace brings together all services on the Open Platform for AI Search. You can view service details without logging in. In the Experience Center, you can use a visual interface to try core capabilities such as document parsing, multimodal embedding and reranking, object detection, text embedding, and video parsing to quickly assess whether the services meet your business needs.

Service overview

Agentic search services

Service category	Description
Agentic Memory	Agentic Memory: Stores and retrieves long-term, short-term, and contextual memory for agents. It manages two types of data: Memory, such as user preferences and interests, and Skill, which contains reusable execution logic. It combines BM25 and vector retrieval in hybrid search to recall memory efficiently and accurately.
Large language models	Qwen3-235B-A22B: A new generation of Qwen large language models. Through extensive training, Qwen3 has achieved breakthroughs in reasoning, instruction following, agent capabilities, and multilingual support. It supports more than 100 languages and dialects, with strong multilingual understanding, reasoning, and generation capabilities. OpenSearch-Qwen-Turbo: A model based on qwen-turbo and fine-tuned through supervised learning to enhance retrieval augmentation and reduce harmful output. Qwen-Turbo: The fastest model in the Qwen series, with very low costs. It is suitable for simple tasks. For more information, see Recommended models. Qwen-Plus: A balanced model whose reasoning performance, cost, and speed fall between those of Qwen-Max and Qwen-Turbo. It is suitable for moderately complex tasks. For more information, see Recommended models. Qwen-Max: The best-performing model in the Qwen series, suitable for complex tasks that involve multiple steps. For more information, see Recommended models. DeepSeek-R1: A large language model that specializes in complex reasoning tasks and performs well in understanding complex instructions and producing accurate results. DeepSeek-V3: A mixture-of-experts (MoE) model that excels at long-text processing, coding, mathematics, encyclopedic knowledge, and Chinese language tasks. DeepSeek-R1-distill-qwen-7b: A model fine-tuned from Qwen-7B through knowledge distillation using training samples generated by DeepSeek-R1. DeepSeek-R1-distill-qwen-14b: A model fine-tuned from Qwen-14B through knowledge distillation using training samples generated by DeepSeek-R1. DeepSeek-V4-Pro: A flagship MoE large language model with 1.6 trillion total parameters and 49 billion active parameters. It natively supports contexts of up to millions of tokens. Trained on large amounts of high-quality data, it provides advanced mathematical reasoning, complex reasoning, professional coding, and in-depth analysis of long texts. It is suitable for demanding scenarios such as advanced scientific research, complex office tasks, and sophisticated intelligent agents. DeepSeek-V4-Flash: An efficient, lightweight MoE model with 284 billion total parameters and 13 billion active parameters. It natively supports contexts of up to millions of tokens. With fast inference, low latency, low invocation costs, and balanced capabilities, it is designed for high-concurrency, lightweight tasks such as everyday conversations, content creation, basic RAG, and batch copywriting.
Internet search	If your private knowledge base cannot answer a query, you can extend the search to the Internet to retrieve additional information. This supplements the private knowledge base and helps a large language model provide richer answers.
Query analysis	General query analysis: Uses large language models and natural language processing (NLP) to recognize intent, expand similar questions, and convert natural language to SQL (NL2SQL) for user queries. This improves retrieval and question answering in retrieval-augmented generation (RAG) scenarios.
Reranking services	BGE reranking model: A document reranking service based on BGE that scores and ranks documents by their semantic relevance to a query. It supports Chinese and English, with a maximum combined input length of 512 tokens for the query and document. OpenSearch reranking model: A high-quality reranking service developed by OpenSearch and trained on datasets from multiple industries. It ranks documents by their semantic relevance to a query. It supports Chinese and English, with a maximum combined input length of 512 tokens for the query and document. Qwen3 reranking service-0.6B: A Qwen3 document reranking service with 0.6 billion parameters. It supports more than 100 languages, with a maximum combined input length of 32k tokens for the query and document. Text reranking-ops-bge-reranker-v2: A high-performance text reranking model that performs fine-grained reranking of retrieval results to improve relevance in scenarios such as RAG.

Text and document processing services

Service category	Description
Jina AI Reader	A web content extraction service designed for large language models. It converts any URL into LLM-friendly plain text, automatically removes distractions such as advertisements and navigation, and accurately extracts the main content of a web page.
Document content parsing	Document content parsing service 001: A general-purpose document parsing service that extracts logical structures, such as titles and paragraphs, from unstructured documents containing text, tables, and images, and returns the results in a structured format. Document content parsing service 002: Parses various unstructured document formats, including PDFs and images. It excels at recognizing complex elements such as tables, formulas, and charts, and provides fast inference.
Text embedding	OpenSearch text embedding service-001: A text embedding service that supports more than 40 languages. The maximum input length is 300 tokens, and the output vector has 1,536 dimensions. OpenSearch general text embedding service-002: A text embedding service that supports more than 100 languages. The maximum input length is 8,192 tokens, and the output vector has 1,024 dimensions. OpenSearch text embedding service-Chinese-001: A Chinese text embedding service. The maximum input length is 1,024 tokens, and the output vector has 768 dimensions. OpenSearch text embedding service-English-001: An English text embedding service. The maximum input length is 512 tokens, and the output vector has 768 dimensions. GTE text embedding-Multilingual-Base: A text embedding service that supports more than 70 languages. The maximum input length is 8,192, and the output vector has 768 dimensions. Qwen3 text embedding-0.6B: A Qwen3 text embedding service with 0.6 billion parameters that supports more than 100 languages. The maximum input length is 32k, and the output vector has 1,024 dimensions. Text embedding-ops-dmeta-text-embedding-zh: A text embedding model optimized for Chinese. It converts Chinese text into high-quality semantic vectors for tasks such as semantic retrieval and text similarity calculation.
Sparse text embedding	OpenSearch sparse text embedding service: Converts text into sparse vectors. Sparse vectors require less storage and are commonly used to represent keywords and term frequencies. They can be combined with dense vectors for hybrid search to improve retrieval results. The service supports more than 100 languages, with a maximum input length of 8,192 tokens.
Document chunking	A general-purpose text chunking service that splits structured data in HTML, Markdown, and TXT formats based on document paragraphs, text semantics, or specified rules. It also extracts code, images, and tables from documents in rich text format.
Vector dimensionality reduction	Vector dimensionality reduction-embedding-dim-reduction: A vector model tuning service that uses custom training to reduce the dimensionality of high-dimensional vectors. It reduces storage costs and improves cost efficiency without significantly affecting retrieval performance.

Multimodal processing services

Service category	Description
Multimodal embedding	M2-Encoder multimodal embedding model: A bilingual Chinese and English multimodal service trained from BM-6B on 6 billion image-text pairs, including 3 billion Chinese pairs and 3 billion English pairs. It supports cross-modal image-text retrieval, including text-to-image and image-to-text search, as well as image classification. M2-Encoder-Large multimodal embedding model: Uses the same architecture as M2-Encoder, with its parameter count increased to 1 billion. This provides stronger representations and better performance on multimodal tasks. GME multimodal embedding-Qwen2-VL-2B: A multimodal embedding service trained on Qwen2-VL multimodal large language models (MLLMs). It supports both single-modal and combined multimodal inputs to efficiently process text, images, and combinations of these data types. Multimodal embedding-ops-mm-embedding-v1-2b: Developed by the Alibaba Cloud OpenSearch AI team and fine-tuned from Qwen2-VL 2B Instruct. It supports single-modal and combined multimodal inputs, including text, images, and video, for cross-modal retrieval and understanding tasks. Multimodal embedding-ops-mm-embedding-v1-7b: Provides the same capabilities as the 2B version. It is fine-tuned from Qwen2-VL 7B Instruct and has more parameters and stronger representational capabilities. E-commerce multimodal embedding-ops-mm-embedding-ecom-001: A custom multimodal embedding model for e-commerce developed by the Alibaba Cloud OpenSearch AI team. It supports image-to-image search and using an image to search image-text content for cross-modal retrieval and understanding in e-commerce scenarios. Face multimodal embedding-ops-mm-embedding-face-001: Uses an advanced face embedding model for face retrieval. It supports one or more input images and encodes facial information into high-dimensional semantic vectors for efficient, accurate face search and comparison.
Multimodal reranking	An image relevance ranking service that reranks retrieval results by relevance in RAG and multimodal search scenarios, improving retrieval accuracy and the accuracy of LLM-generated output.
Video summarization	A video summarization service that understands a specified video segment and uses a large language model to generate a summary, title, and tags.
Video annotation	A video annotation service that analyzes a specified video frame by frame and uses a large language model to generate object detection bounding boxes, keypoints, and semantic scene annotations.
Video splitting	A video splitting service that understands and analyzes a video, extracts keyframes, and splits the video into corresponding segments.
Video frame extraction	Video frame extraction service 001: Extracts keyframes from videos. Combined with multimodal embedding or image parsing, it enables cross-modal retrieval.
Speech recognition	Speech recognition service 001: Provides speech-to-text capabilities that quickly convert speech in video or audio into structured text. It supports multiple languages.
Object detection	Object detection service-ops-object-detect-001: Automatically locates and identifies the main objects in images or videos. It supports single-object and multiple-object detection for applications such as intelligent surveillance, autonomous driving, and image retrieval. Face detection service-ops-object-detect-face-001: Automatically locates and detects faces in images. It supports detecting multiple faces for applications such as intelligent surveillance and image retrieval. Fashion object detection 001-ops-object-detect-fashion-001: A fashion object detection model developed by the Alibaba Cloud OpenSearch team. It detects more than 20 categories of fashion objects and returns their categories, bounding boxes, and confidence scores. It extracts objects from complex scenes for use in downstream tasks, reducing interference from image backgrounds or other objects in downstream embedding and reranking models.
Image content parsing	Image content understanding service: Uses a multimodal large language model to parse and understand image content and recognize text. The resulting text can be used for image retrieval and question answering.
Image content parsing	Image text recognition service: Uses optical character recognition (OCR) to recognize text in images. The recognized text can be used for image retrieval and question answering.

Features

The Experience Center provides the following services:

Service group	Service category	Description
Agentic search services	Agentic Memory Provides long-term, short-term, and contextual memory storage and retrieval for Agentic AI agents and intelligent search services. It supports storing, querying, updating, and forgetting various types of search data, including personalized memory, agent memory, search memory, and Skill management data. It combines BM25 and vector retrieval with hybrid search and multiple recall paths to retrieve memory efficiently and accurately.	Agentic Memory: Stores and manages two types of data. Memory stores information such as users' hobbies, interests, and preferences. Skill stores reusable execution logic and skills.
	Large language models	Qwen3-235B-A22B: A new generation of Qwen large language models. Through extensive training, Qwen3 has achieved breakthroughs in reasoning, instruction following, agent capabilities, and multilingual support. It supports more than 100 languages and dialects, with strong multilingual understanding, reasoning, and generation capabilities. OpenSearch-Qwen-Turbo: A model based on qwen-turbo and fine-tuned through supervised learning to enhance retrieval augmentation and reduce harmful output. Qwen-Turbo: The fastest model in the Qwen series, with very low costs. It is suitable for simple tasks. For more information, see Recommended models. Qwen-Plus: A balanced model whose reasoning performance, cost, and speed fall between those of Qwen-Max and Qwen-Turbo. It is suitable for moderately complex tasks. For more information, see Recommended models. Qwen-Max: The best-performing model in the Qwen series, suitable for complex tasks that involve multiple steps. For more information, see Recommended models. DeepSeek-R1: A large language model that specializes in complex reasoning tasks and performs well in understanding complex instructions and producing accurate results. DeepSeek-V3: A mixture-of-experts (MoE) model that excels at long-text processing, coding, mathematics, encyclopedic knowledge, and Chinese language tasks. DeepSeek-R1-distill-qwen-7b: A model fine-tuned from Qwen-7B through knowledge distillation using training samples generated by DeepSeek-R1. DeepSeek-R1-distill-qwen-14b: A model fine-tuned from Qwen-14B through knowledge distillation using training samples generated by DeepSeek-R1. DeepSeek-V4-Pro: A flagship MoE large language model with 1.6 trillion total parameters and 49 billion active parameters. It natively supports contexts of up to millions of tokens. Trained on large amounts of high-quality data, it provides advanced mathematical reasoning, complex reasoning, professional coding, and in-depth analysis of long texts. It is suitable for demanding scenarios such as advanced scientific research, complex office tasks, and sophisticated intelligent agents. DeepSeek-V4-Flash: An efficient, lightweight MoE model with 284 billion total parameters and 13 billion active parameters. It natively supports contexts of up to millions of tokens. With fast inference, low latency, low invocation costs, and balanced capabilities, it is designed for high-concurrency, lightweight tasks such as everyday conversations, content creation, basic RAG, and batch copywriting.
	Internet search	If your private knowledge base cannot answer a query, you can extend the search to the Internet to retrieve additional information. This supplements the private knowledge base and helps a large language model provide richer answers.
	Query analysis Provides query content analysis based on large language models and NLP, including intent recognition, similar question expansion, and NL2SQL processing for user queries. This improves retrieval and question answering in RAG scenarios.	General query analysis: Uses a large language model to understand the intent of user queries and expand similar questions.
	Reranking services	BGE reranking model: A BGE-based document scoring service that scores documents by their relevance to a query, sorts them in descending order of score, and returns the corresponding scores. It supports Chinese and English, with a maximum combined input length of 512 tokens for the query and document. OpenSearch reranking model: A high-quality reranking service developed by OpenSearch and trained on datasets from multiple industries. It sorts documents in descending order of semantic relevance to a query. It supports Chinese and English, with a maximum combined input length of 512 tokens for the query and document. Qwen3 reranking service-0.6B: A Qwen3 document reranking service with 0.6 billion parameters. It supports more than 100 languages, with a maximum combined input length of 32k tokens for the query and document. Text reranking-ops-bge-reranker-v2: A high-performance text reranking model that performs fine-grained reranking of retrieval results to improve relevance in scenarios such as RAG.
	Reranking services
Text and document processing services	Jina AI Reader	A web content extraction service designed for large language models. It converts any URL into LLM-friendly plain text, automatically removes distractions such as advertisements and navigation, and accurately extracts the main content of a web page.
	Document content parsing	Document content parsing service 001: A general-purpose document parsing service that extracts logical structures, such as titles and paragraphs, from unstructured documents containing text, tables, and images, and returns the results in a structured format. Document content parsing service 002: Parses various unstructured document formats, including PDFs and images. It excels at recognizing complex elements such as tables, formulas, and charts, and provides fast inference.
	Text embedding	OpenSearch text embedding service-001: A text embedding service that supports more than 40 languages. The maximum input length is 300 tokens, and the output vector has 1,536 dimensions. OpenSearch general text embedding service-002: A text embedding service that supports more than 100 languages. The maximum input length is 8,192 tokens, and the output vector has 1,024 dimensions. OpenSearch text embedding service-Chinese-001: A Chinese text embedding service. The maximum input length is 1,024 tokens, and the output vector has 768 dimensions. OpenSearch text embedding service-English-001: An English text embedding service. The maximum input length is 512 tokens, and the output vector has 768 dimensions. GTE text embedding-Multilingual-Base: A text embedding service that supports more than 70 languages. The maximum input length is 8,192, and the output vector has 768 dimensions. Qwen3 text embedding-0.6B: A Qwen3 text embedding service with 0.6 billion parameters that supports more than 100 languages. The maximum input length is 32k, and the output vector has 1,024 dimensions. Text embedding-ops-dmeta-text-embedding-zh: A text embedding model optimized for Chinese. It converts Chinese text into high-quality semantic vectors for tasks such as semantic retrieval and text similarity calculation.
	Sparse text embedding Converts text into sparse vectors. Sparse vectors require less storage and are commonly used to represent keywords and term frequencies. They can be combined with dense vectors for hybrid search to improve retrieval results.	OpenSearch sparse text embedding service: A text embedding service that supports more than 100 languages, with a maximum input length of 8,192 tokens.
	Document chunking	A general-purpose text chunking service that splits structured data in HTML, Markdown, and TXT formats based on document paragraphs, text semantics, or specified rules. It also extracts code, images, and tables from documents in rich text format.
	Vector dimensionality reduction	Vector dimensionality reduction-embedding-dim-reduction: A vector model tuning service that uses custom training for models such as dimensionality reduction models. It reduces the dimensionality of high-dimensional vectors with limited impact on retrieval performance, improving cost efficiency.
Multimodal processing services	Multimodal embedding	M2-Encoder multimodal embedding model: A bilingual Chinese and English multimodal service trained from BM-6B on 6 billion image-text pairs, including 3 billion Chinese pairs and 3 billion English pairs. It supports cross-modal image-text retrieval, including text-to-image and image-to-text search, as well as image classification. M2-Encoder-Large multimodal embedding model: A bilingual Chinese and English multimodal service. Compared with M2-Encoder, it has more parameters, reaching 1 billion, to provide stronger representations and better performance on multimodal tasks. GME multimodal embedding-Qwen2-VL-2B: A multimodal embedding service trained on Qwen2-VL multimodal large language models (MLLMs). It supports both single-modal and combined multimodal inputs to efficiently process text, images, and combinations of these data types. Multimodal embedding-ops-mm-embedding-v1-2b: A multimodal embedding model developed by the Alibaba Cloud OpenSearch AI team and fine-tuned from Qwen2-VL 2B Instruct. It supports single-modal and combined multimodal inputs, including text, images, and video, and encodes them into semantic vectors for cross-modal retrieval and understanding tasks. Multimodal embedding-ops-mm-embedding-v1-7b: A multimodal embedding model developed by the Alibaba Cloud OpenSearch AI team and fine-tuned from Qwen2-VL 7B Instruct. It supports single-modal and combined multimodal inputs, including text, images, and video, and encodes them into semantic vectors for cross-modal retrieval and understanding tasks. E-commerce multimodal embedding-ops-mm-embedding-ecom-001: A custom multimodal embedding model for e-commerce developed by the Alibaba Cloud OpenSearch AI team. It supports image-to-image search and using an image to search image-text content for cross-modal retrieval and understanding in e-commerce scenarios. Face multimodal embedding-ops-mm-embedding-face-001: Uses an advanced face embedding model for face retrieval. It supports one or more input images and encodes facial information into high-dimensional semantic vectors for efficient, accurate face search and comparison.
	Multimodal reranking	An image relevance ranking service that finds and returns content in order of relevance in RAG and multimodal search scenarios. It improves retrieval accuracy and the accuracy of LLM-generated output.
	Video summarization	A video summarization service that understands a specified video segment and uses a large language model to generate a summary, title, and tags.
	Video annotation	A video annotation service that analyzes a specified video frame by frame and uses a large language model to generate object detection bounding boxes, keypoints, and semantic scene annotations.
	Video splitting	A video splitting service that understands and analyzes a video, extracts keyframes, and splits the video into corresponding segments.
	Video frame extraction	Video frame extraction service 001: Extracts keyframes from videos. Combined with multimodal embedding or image parsing, it enables cross-modal retrieval.
	Speech recognition	Speech recognition service 001: Provides speech-to-text capabilities that quickly convert speech in video or audio into structured text. It supports multiple languages.
	Object detection	Object detection service-ops-object-detect-001: Automatically locates and identifies the main objects in images or videos. It supports single-object and multiple-object detection for applications such as intelligent surveillance, autonomous driving, and image retrieval. Face detection service-ops-object-detect-face-001: Automatically locates and detects faces in images. It supports detecting multiple faces for applications such as intelligent surveillance and image retrieval. Fashion object detection 001-ops-object-detect-fashion-001: A fashion object detection model developed by the Alibaba Cloud OpenSearch team. It detects more than 20 categories of fashion objects and returns their categories, bounding boxes, and confidence scores. It extracts objects from complex scenes for use in downstream tasks, reducing interference from image backgrounds or other objects in downstream embedding and reranking models.
	Image content parsing	Image content understanding service: Uses a multimodal large language model to parse and understand image content and recognize text. The resulting text can be used for image retrieval and question answering.
	Image content parsing	Image text recognition service: Uses optical character recognition (OCR) to recognize text in images. The recognized text can be used for image retrieval and question answering.

Try services

This section uses document parsing and multimodal embedding as examples to show how to quickly try services in the Experience Center, view results, and obtain code for calling the services.

Document parsing

Log on to the Open Platform for AI Search console.
In the left-side navigation pane, select Experience Center.
For Service Category, select Document/Image Parsing (document-analyze), and then select a service from Experience Services.
Use the system-provided Sample data, or upload your own data by using Manage data. Supported file formats: TXT, PDF, HTML, DOC, DOCX, PPT, and PPTX. The maximum file size is 20 MB.
- File: Upload local files. Files are automatically deleted after 7 days. The platform does not store your data long-term.
- URL: Enter the file URLs and corresponding file types. You can enter multiple URLs, with one URL per line.
  
  Note
  Selecting an incorrect data format causes document parsing to fail. Select the correct file type for your data.
  
  Important
  Use the web link import feature in compliance with applicable laws and regulations, follow the rules of the target platform, and protect the legitimate rights and interests of rights holders. You are independently responsible for your actions. As the tool provider, the Open Platform for AI Search is not liable for your parsing or downloading activities.
If you use your own data, select a previously uploaded file or URL from the drop-down list.
Click Get Results to call the service and parse the document.
- Results: Displays the parsing progress and results.
- Result source code: View the response code. Click Copy Code or Download File to save the code to your computer.
- Sample code: View and download the Sample code for calling the text content parsing service.

Multimodal embedding

Log on to the Open Platform for AI Search console.
In the left-side navigation pane, select Experience Center.
For Service Category, select Multimodal Vector (multi-modal-embedding), and then select a service from Experience Services. Select Text, Image, or Text + Image.

Note
Local images uploaded for embedding are automatically deleted after 7 days. The platform does not store your data long-term.
Click Get Results to obtain the multimodal embedding results.
- Results: Displays the embedding results.
- Result source code: View the response code. Click Copy Code or Download File to save the code to your computer.
- Sample code: View and download the Sample code for calling the text embedding service.