OpenSearch LLM-Based Conversational Search Edition is a fully managed Retrieval-Augmented Generation (RAG) platform that combines unstructured data processing, vector models, text and vector retrieval, and large language models (LLMs). Upload your business data and get a working RAG system — including conversational answers, reference links, and images — in minutes.
Service architecture
What you get
When you create an OpenSearch LLM-Based Conversational Search Edition instance, you get:
Multimodal RAG — process images alongside text using optical character recognition (OCR) and LLMs to build a multimodal knowledge base with diverse output formats.
RAG performance evaluation — run end-to-end evaluations across different model and parameter configurations to compare RAG quality and select the best setup.
Built-in model suite — vector, reranking, and LLMs are included out of the box; train exclusive models on your own business data (available in Professional Edition).
Real-time data updates — build incremental vector indexes and sync data changes as they happen, keeping search results current.
Table Q&A — convert natural language into SQL queries (NL2SQL) to search enterprise structured databases conversationally.
Zero-ops deployment — fully managed Model as a Service (MaaS); no infrastructure to deploy or maintain.
Over 95% RAG accuracy — built-in model capabilities that have topped industry rankings multiple times.
Flexible integration — connect via API/SDK, embed an API key into open-source LLM frameworks, or enable zero-code access for DingTalk and Lark bots.
Enterprise controls — document-level permission isolation and real-time incremental data updates for enterprise knowledge management.
How it works
Upload business data (documents, databases, or web content) through the console.
The service chunks, vectorizes, and indexes the data automatically.
Submit a natural language query via the conversational interface or API.
The service retrieves relevant content, reranks results, and generates an LLM answer with reference links and images.
Evaluate RAG quality and fine-tune prompts, model parameters, or search ranking as needed.
Use cases
Intelligent customer service
Add pre-sales and after-sales support to apps, miniapps, and websites. The service classifies user intent and returns the appropriate answer, including:
Fixed Q&A pairs maintained through manual intervention.
Multimodal responses such as images and videos.
Order and logistics lookups powered by NL2SQL database queries.
Enterprise knowledge base
Deploy an internal knowledge base on portal sites or chat platforms for employee self-service and quick navigation:
Document-level permission isolation keeps sensitive content within the right teams.
Real-time index updates reflect the latest internal data immediately.
Unstructured data in various formats is parsed and understood automatically.
E-commerce shopping guide
Add an intelligent shopping assistant to existing search boxes or customer service chat in retail apps and websites:
Multimodal understanding lets users search with product images as well as text.
Results include direct product links for one-click access to the target item.
Product ranking adjusts based on operational priorities.
Content and community summary
Augment the search box in content platforms and community sites with AI-generated summaries:
Import webpage and website content quickly at scale.
OCR and LLMs extract meaning from images embedded in content.
Responses cite reference links back to the original documents.
Table Q&A
Search structured financial and business data using plain language:
NL2SQL translates natural language queries into SQL and summarizes results via LLM.
Custom table structures are supported.
MaxCompute data sources sync automatically.
Choose a version
OpenSearch LLM-Based Conversational Search Edition offers two versions.
| Standard Edition | Professional Edition | |
|---|---|---|
| Customized model training | Not supported | Supports SFT based on your own business data |
| LLM selection | Qwen series, open-source models, and external models | Qwen series, open-source models, external models, and customized models |
| Throughput limit | Maximum 10 queries per second (QPS) | No throttling, limited only by purchased GPU capacity |
| Billing | Pay-as-you-go based on computing resources consumed per invocation | Billed by purchased GPU specifications; no additional per-invocation fees |
| Best for | General intelligent customer service, enterprise knowledge bases, and e-commerce shopping guides | Scenarios requiring exclusive LLMs trained on proprietary business data |