AI Search Open Platform provides component-based algorithm services for intelligent search and retrieval-augmented generation (RAG) scenarios. It covers the full pipeline — document parsing, chunking, text embedding, query analysis, retrieval, sorting, performance evaluation, and large language model (LLM)-based generation — with each stage exposed as an independent, callable service. Select only the components your use case requires.
Activation is free. New Alibaba Cloud accounts receive 10 free service calls, shared across all RAM users in the account. After the free calls are used, charges apply based on actual LLM service call usage. Activate Now
Use cases
RAG and conversational search
Build knowledge bases, intelligent customer service systems, and conversational search applications. Typical queries: "What are the steps to configure X?" or "Summarize the key points in this document."
Intelligent customer service
Conversational search
Knowledge graph enhancement
Personalized recommendation
For a development example, see Build a RAG-based conversational search application.
Multi-modal search
Search across text and image content using semantic understanding and visual recognition. Typical queries: image-based lookups, cross-modal content discovery, and OCR-powered document search.
E-commerce and retail
News content
Gaming
Healthcare
Finance
For a development example, see Multi-modal search business development.
How it works
AI Search Open Platform processes data through a modular pipeline. Each stage is an independent, callable service:
Parse documents — Extract text, tables, images, and code from PDF, DOC, HTML, TXT, and other formats. Minute-level parsing speed.
Chunk content — Split documents by semantics and paragraph structure. The resulting chunk tree supports context completion during retrieval.
Embed text — Convert chunks into dense or sparse vectors using multilingual embedding models. Combine both for hybrid search.
Analyze queries — Understand user intent, expand similar questions, or convert natural language into SQL using LLM and natural language processing (NLP) capabilities.
Retrieve results — Run vector and text retrieval against indexed data.
Sort and rank — Apply query- and document-level sorting to surface the most relevant content.
Generate responses — Use built-in LLMs to produce answers grounded in retrieved content.
Features
Document content parsing
Parses PDF, DOC, HTML, TXT, and other formats at minute-level speed. Distinguishes layouts, extracts logical structures (titles, paragraphs), and identifies content elements including text, tables, images, and code. Removes headers and footers and detects superscripts and subscripts. Output is structured.
Image content parsing
Parses image content — such as architecture diagrams and analytical charts — using multi-modal LLMs. Supports optical character recognition (OCR) to extract text from images for use in image retrieval and image-based Q&A.
Document chunking
Splits documents by semantics, paragraph structure, and configurable rules. Generates a chunk tree for context completion during retrieval.
Multilingual embedding models
Text embedding: Converts text to dense vectors. Multiple models support different languages, input lengths, and output dimensions. Use for search, text classification, and relevance comparison.
Sparse embedding: Converts text to sparse vectors that use less storage. Captures keywords and high-frequency terms. Combine with dense vectors for hybrid search to improve retrieval performance.
Embedding model tuning: Train a dimensionality reduction model to reduce vector dimensions without significantly affecting retrieval quality.
Query analysis
Analyzes queries using LLMs and NLP to understand user intent, expand similar questions, and convert natural language into SQL. Improves the accuracy of conversational search in RAG scenarios.
Search engine
Provides vector and text retrieval engines. Store vectors and text, build indexes, and run online vector and text retrieval. Use the engines with AI Search Open Platform APIs to process and retrieve data end to end.
Sorting
Applies query- and document-level sorting in RAG and search scenarios to surface the most relevant content. Improves retrieval accuracy and the quality of LLM-generated responses.
LLM-based text generation
Provides a range of models, including Qwen3-235B-A22B, QwQ, all DeepSeek models (DeepSeek R1, V3, and 7B/14B distilled variants), and the Qwen series (Qwen-Turbo, Qwen-Plus, and Qwen-Max). The built-in OpenSearch-Qwen-Turbo model is based on qwen-turbo and enhanced in RAG capabilities after supervised fine-tuning to reduce the hallucination rate.
Benefits
End-to-end pipeline, no assembly required — All services are available immediately after activation. Rather than integrating separate tools for each stage of the search or RAG pipeline, you get a single platform covering parsing, embedding, retrieval, ranking, and generation.
Call what you need, skip what you don't — Developers, enterprise customers, and ISVs (independent software vendors) can call individual API operations or use SDKs to integrate any subset of services. The modular design means your architecture only carries the components it uses.
Search-optimized models out of the box — AI Search Open Platform lets you train dedicated AI search models on top of leading foundation models. The built-in OpenSearch-Qwen-Turbo model is fine-tuned specifically for RAG, reducing hallucinations compared to a general-purpose model.
Built on proven search patterns — The pipeline design reflects years of experience in intelligent search and RAG. Best practices are built in, so you spend less time learning what works and more time building.