This topic describes the feature updates and corresponding documentation of AI Search Open Platform.
2025
June
Type | Feature | Description | Release date | Reference |
New | Speech recognition | AI Search Open Platform supports calling speech recognition services through APIs, which can quickly convert speech content in videos or audio into structured text. This feature can be used for meeting records, video retrieval, online customer service, and other scenarios. | 2025-06-26 | |
New | Video snapshot | AI Search Open Platform supports calling video snapshot services through APIs, which can extract keyframes from videos. Combined with optical character recognition (OCR), image parsing, or multimodal embedding services, it enables deep parsing and structured processing of video content. | 2025-06-26 | |
Update | ops-qwen3-embedding-0.6b added to text embedding | Multilingual (100+) text embedding service of the Qwen3 series, with a maximum input length of 32k and customizable output vector dimensions from 32 to 1024, with 0.6B parameters. | 2025-06-26 | |
Update | ops-qwen3-reranker-0.6b added to re-rank service | Document re-ranking service of the Qwen3 series, which supports 100+ languages with a maximum input token length of 32k (Query + doc length) and 0.6B parameters. | 2025-06-26 | |
Update | ops-gme-qwen2-vl-2b-instruct added to multimodal embedding service | A multimodal embedding service trained based on the Qwen2-VL multimodal large language model (MLLM). It supports single-modal and multimodal combination inputs and can efficiently process text, images, and combined data types. | 2025-06-26 | |
New | Multimodal embedding | Multimodal embedding service trained based on the Qwen2-VL multimodal large language model (MLLM). It supports single-modal and multimodal combination inputs and can efficiently process text, images, and combined data types. | 2025-06-4 |
April
Type | Feature | Description | Release date | Reference |
Update | Qwen3-235B-A22B added to LLM service | This model is a new-generation Qwen series LLM that is extensively trained. Qwen3 has made significant breakthroughs in inference, instruction following, agent capability, and multi-language support, can support more than 100 languages and dialects, and has powerful multi-language understanding, inference, and generation capabilities. | 2025-04-29 |
March
Type | Feature | Description | Release date | Reference |
New | Web search | The web search feature is added. You can call the web search API independently or use the web search feature during LLM-based conversational research. | 2025-03-20 | |
Update | QwQ deep thinking model | QwQ reasoning model trained based on the Qwen2.5-32B model. The model's reasoning capability is greatly improved through reinforcement learning. The performance of this model in math and coding (AIME 24/25 and LiveCodeBench) and some of its general performance indicators, such as IFEval and LiveBench, have reached the level of the DeepSeek-R1 full version. | 2025-03-20 | |
Update | Model deployment | Model source is ModelScope Model types: multimodal embedding models and re-ranking models. | 2025-03-13 | |
New | Service development | AI Search Open Platform is integrated with Data Science Workshop (DSW) of Platform for AI (PAI), allowing developers to complete service development and execution using Notebook in DSW. | 2025-03-13 |
February
Type | Feature | Description | Release date | Reference |
New | DeepSeek models | All DeepSeek models (including DeepSeek R1/V3 and 7B/14B distilled models) are supported by AI Search Open Platform. In addition, AI Search Open Platform is compatible with the OpenAI API for calling. | 2025-02-14 |
January
Type | Feature | Description | Release date | Reference |
New | Model deployment | Models are independently deployed in AI Search Open Platform, which provides inference services with higher concurrency and lower latency. | 2025-01-07 |
2024
October
Type | Feature | Description | Release date | Reference |
New | Vector dimensionality reduction service | Custom training of vector dimensionality reduction models is supported based on user-provided vector data. | 2024-10-23 |
September
Type | Feature | Description | Release date | Reference |
New | Re-ranking model | The self-developed re-ranking model is added, which is trained with multi-industry datasets and supports multiple languages. In eight retrieval tasks on Chinese MTEB evaluation, the re-ranking model outperforms the open-source bge-rerank series models. | 2024-09-12 | |
New | NL2SQL added in the query analysis service | The natural language to SQL capability is added in the query analysis service, supporting NL2SQL service configuration and calling. | 2024-09-12 | |
New | Multi-modal data processing scenario | The multi-modal data processing scenario is added to the scene center, which supports text and image parsing and embedding. | 2024-09-12 | |
New | Llamaindex code architecture added to the RAG scenario | The Llamaindex code architecture is added to the scenario center, allowing you to build a RAG-based conversational search application. | 2024-09-12 |
August
Type | Feature | Description | Release date | Reference |
Optimized | OCR optimization in image parsing | The performance is improved by about 40% with OCR optimization in image parsing. | 2024-08 |
July
Type | Feature | Description | Release date | Reference |
New | Evaluation management | The effect evaluation module is added to evaluate the effectiveness of the RAG pipeline. After you upload an evaluation dataset, you can obtain evaluation results based on an LLM. | 2024-07-08 | |
New | Data upload in the experience center | Self-managed data can be uploaded to the experience center for service experience. | 2024-07-08 | |
New | Compatible with the OpenAI SDK | AI Search Open Platform is compatible with OpenAI interfaces, allowing developers to call services provided by AI Search Open Platform using the OpenAI SDK. | 2024-07-08 | |
New | Langchain code architecture added to the RAG scenario | The Langchain code architecture is added to the scenario center, allowing you to build a RAG-based conversational search application. | 2024-07-08 | |
New | Query analysis service | The query analysis service is added for queries based on LLMs and NLP capabilities to understand the intent of users, extend similar questions, and convert questions in natural languages into SQL statements. This improves the performance of conversational search in RAG scenarios. | 2024-07-08 | |
New | Image parsing service | The image parsing service is added, which lets you parse the content of an image based on multimodal LLMs. You can also parse the text in the image and use the parsed text for image retrieval and conversational search. | 2024-07-08 | |
New | Query analysis service experience | The query analysis service is added to the experience center. You can experience the service to provide intents and similar queries based on historical messages and queries. | 2024-07-30 | |
New | Image parsing in RAG scenarios | Image parsing in documents is supported in RAG scenarios and the query analysis service is supplemented. This lets you perform conversational search based on the rewritten queries. | 2024-07-30 | |
New | VPC access regions | VPC access to services is supported in the China (Shanghai), China (Hangzhou), China (Shenzhen), China (Beijing), China (Qingdao), and China (Zhangjiakou) regions. | 2024-07-30 |
June
Type | Feature | Description | Release date | Reference |
New | RAG-based conversational search application | A RAG-based solution is provided for conversational search in knowledge bases. The solution consists of the following modules: data preprocessing, data retrieval, and output generation. | 2024-06-18 |
May
Type | Feature | Description | Release date | Reference |
New | LLM | LLMs are added, including Qwen and fine-tuned RAG-specific LLMs. | 2024-05-21 | |
New | Re-ranking service | The re-ranking service is added, which provides the general document scoring capability. The ranking service sorts the docs in descending order based on the relevance between the query and document content and outputs the scoring result. | 2024-05-21 | |
New | Sparse text embedding service | The sparse text embedding service is added, which converts text data to sparse vectors. | 2024-05-21 | |
New | Text embedding service | The text embedding service is added, which converts text data to dense vectors. | 2024-05-21 | |
New | Document chunking service | The text chunking service is added. You can use this service to split structured data in HTML, MARKDOWN, and TXT formats based on paragraph structures, semantics, and specific rules. You can also extract code, images, and tables from rich texts. | 2024-05-21 | |
New | Document parsing service | The document parsing service is added, which lets you extract logical hierarchical structures and the information from an unstructured document and generate the output content in a structured format. The hierarchical structures include titles and paragraphs, and the information includes text, tables, and images. | 2024-05-21 | |
New | Service experience | The console service experience capability is added, which supports text parsing, chunking, embedding, re-ranking, LLM, and other service experiences. | 2024-05-21 |