Introducing Intelligent Search Development
Download
フォーカスモード
フォントサイズ
The new Intelligent Search Development feature released by Tencent Cloud ES further enhances its support for hybrid full-text and vector search capabilities. It provides comprehensive support and optimization across the entire workflow, from atomic capabilities like raw document parsing and vectorization to query performance, hybrid ranking efficiency, and search result accuracy, unlocking greater potential for search applications. Building on this foundation, it can also be seamlessly integrated with large language models like DeepSeek. This enables enterprises to more efficiently and flexibly build RAG applications, such as knowledge Q&A systems, and allows for fine-tuning at critical stages. The overall process is as follows:
Atomic Services
Enabling Atomic Services:
API calls are supported only after atomic services are activated. Go to the console promptly to activate the service. Specific steps: Click Activate Now and confirm the free activation.
Currently, the following types of atomic services are provided by Intelligent Search Development:
Type | Description | Service | Description |
Document parsing | This service can convert multi-format documents such as PDF, DOC, PPT, XLS, HTML, TXT, and PNG into a standard format, meeting the needs of enterprise knowledge base construction, technical document migration, and structured storage for content platforms. | doc-llm | A premium OCR large language model launched by Tencent Youtu Lab, covering a wide range of scenarios such as papers, books, manuals, exam papers, PPTs, and posters, and meeting the document processing needs of various industries. |
Text chunking | Text chunking is a technique for splitting long texts into shorter segments. It is used to adapt to model input, improve processing efficiency, or facilitate information search, while balancing segment length and semantic coherence. This technique applies to scenarios such as NLP and data analysis. This service supports semantic segmentation. | doc-chunk | Slice based on separators and text length, applicable to texts with strong regularities, and only support Markdown and TXT file input. |
| | doc-tree-chunk | A self-developed document splitting model by Tencent, supporting multiple file types, capable of parsing and deeply understanding information in charts and graphs, and semantically chunking it into text blocks. |
Embedding | A service that converts multimodal data such as text and images into dense or sparse vector forms. It supports multiple text vector models with different languages, input lengths, and dimensions, and can be used for scenarios such as information search, classification, and similarity comparison. | 896 dimensions, maximum 131072 tokens A self-developed multilingual model by WeChat, supporting multiple languages. | |
| | 1024 dimensions, maximum 8194 tokens A classic BGE Embedding model, supporting 100+ languages, and suitable for multilingual search scenarios. | |
| | WeCLIPv2-Base | |
| | WeCLIPv2-Large | 768 dimensions, maximum 72 tokens A self-developed image-text Embedding model by WeChat, which performs excellently in Chinese scenarios and is suitable for scenarios such as text-to-image search and image-to-image search. |
Reranking | In RAG scenarios, more relevant content can be found through the RAG ranking service. Introducing the ranking service can effectively improve the accuracy of search and large model generation while reducing large model costs. | Maximum 8192 tokens per document An upgraded BGE reranking model, featuring a larger text length and supporting multiple languages. | |
Use Cases
Text Search: Leveraging inverted indexes and tokenization optimization, Tencent Cloud ES supports millisecond-level search across billions of data entries. It features capabilities such as spell correction, synonym expansion, and highlight display. Widely applied in scenarios like e-commerce product search and news recommendation, it accurately matches user query intent.
Multimodal Search: It provides vectorization services for multimodal data. By leveraging the hybrid search capabilities of Tencent Cloud ES, it supports joint search across mixed data types such as text and images, enabling scenarios like image-to-image search, text-to-image search, and text-to-video search, thereby breaking through the limitations of a single modality.
Knowledge Base Q&A: Through natural language processing and semantic analysis, it quickly matches user questions with knowledge base content. It then passes the user question and retrieved content to a large language model for summarization and generation. This applies to scenarios such as enterprise FAQs and customer service robots, enabling efficient knowledge retrieval and precise answer delivery.
AI Assistant: It provides real-time knowledge base search and conversational context management for AI assistants. Through search enhancement and multi-turn dialogue support, it improves the response capabilities of virtual assistants, intelligent customer service, and similar applications.
Strengths
Text/Vector Hybrid Search: It features a unique hybrid search capability for both text and vectors, supporting joint search across multimodal data. Through optimizations in inverted indexes and extensions to vector algorithms, it achieves precise matching and enhanced recall in complex scenarios.
Open Atomic Services: They provide atomic services for document parsing, text chunking, Embedding, reranking, LLM generation, and web search. By integrating the ES Inference API, they lower the development barrier for applications such as RAG and AI search.
Flexible AI application building: Highly flexible write/search workflow customization capabilities support calling built-in atomic services and custom services to quickly build scenarios such as intelligent Q&A and AI search.
High Performance and Stability: It supports a hundred-billion-level vector scale and delivers millisecond-level response latency. The self-developed storage optimization saves 70%-90% of resources. By leveraging a distributed architecture and multi-AZ disaster recovery, it ensures stable operation under high-concurrency scenarios.
Authoritative Certification: As the first formulator and certified passer (certified by CAICT) of the end-to-end RAG technical standard, Tencent Cloud ES has, over the past year, facilitated the deployment of top-tier applications such as WeChat Read AI Q&A, ima, and the Tencent Meeting Intelligent Assistant.
フィードバック