Build Chinese Vector Search with Alibaba Cloud AI and Elasticsearch Inference APIs

Sparse Vector

Create a sparse‑embedding inference endpoint:

PUT _inference/sparse_embedding/alibabacloud_ai_search_sparse
{
  "service": "alibabacloud-ai-search",
  "service_settings": {
    "api_key": "<api_key>",
    "service_id": "ops-text-sparse-embedding-001",
    "host": "default-j01.platform-cn-shanghai.opensearch.aliyuncs.com",
    "workspace": "default"
  }
}

Test the endpoint with a Chinese sentence:

POST _inference/alibabacloud_ai_search_sparse
{
  "input": "阿里巴巴（中国）有限公司成立于2007年03月26日，法定代表人蒋芳"
}

The response shows a Unicode‑encoded sparse vector, different from the English‑only ELSER model.

Dense Vector

Create a dense‑embedding inference endpoint:

PUT _inference/text_embedding/alibabacloud_ai_search_embeddings
{
  "service": "alibabacloud-ai-search",
  "service_settings": {
    "api_key": "<api_key>",
    "service_id": "ops-text-embedding-001",
    "host": "default-j01.platform-cn-shanghai.opensearch.aliyuncs.com",
    "workspace": "default"
  }
}

Generate a dense vector:

POST _inference/alibabacloud_ai_search_embeddings
{
  "input": "阿里巴巴（中国）有限公司成立于2007年03月26日，法定代表人蒋芳"
}

Dense vectors are floating‑point arrays that can be scalar‑quantized to save memory and speed up search.

Completion

Create a completion inference endpoint:

PUT _inference/completion/alibabacloud_ai_search_completion
{
  "service": "alibabacloud-ai-search",
  "service_settings": {
    "host": "default-j01.platform-cn-shanghai.opensearch.aliyuncs.com",
    "api_key": "{{API_KEY}}",
    "service_id": "ops-qwen-turbo",
    "workspace": "default"
  }
}

Example request:

POST _inference/completion/alibabacloud_ai_search_completion
{
  "input": "阿里巴巴（中国）有限公司是什么时候成立的?"
}

The large model returns a natural‑language answer based on its training.

Rerank

Create a rerank inference endpoint:

PUT _inference/rerank/alibabacloud_ai_search_rerank
{
  "service": "alibabacloud-ai-search",
  "service_settings": {
    "api_key": "<api_key>",
    "service_id": "ops-bge-reranker-larger",
    "host": "default-j01.platform-cn-shanghai.opensearch.aliyuncs.com",
    "workspace": "default"
  }
}

Provide two candidate documents and a query; the endpoint returns the document with higher relevance.

RAG Application

Store business or private data in Elasticsearch as a vector database. Search the index first, then feed the retrieved documents as context to the completion endpoint to avoid hallucinations.

PUT alibaba_sparse
{
  "mappings": {
    "properties": {
      "inference_field": {
        "type": "semantic_text",
        "inference_id": "alibabacloud_ai_search_sparse"
      }
    }
  }
}

Index two sample documents and run a semantic search:

GET alibaba_sparse/_search
{
  "query": {
    "semantic": {
      "field": "inference_field",
      "query": "阿里云是什么时候成立的？"
    }
  }
}

The result shows the Alibaba Cloud document ranked first.

Creating a Dense Index for RAG

PUT alibaba_dense
{
  "mappings": {
    "properties": {
      "inference_field": {
        "type": "semantic_text",
        "inference_id": "alibabacloud_ai_search_embeddings"
      }
    }
  }
}

After indexing similar documents, semantic searches for "Alibaba 的法人是谁？" correctly retrieve the Alibaba document, demonstrating cross‑language matching.

Automatic Chunking

The semantic_text field automatically splits long texts into chunks for vectorization. Create a large‑text index:

PUT alibaba_dense_large_text
{
  "mappings": {
    "properties": {
      "inference_field": {
        "type": "semantic_text",
        "inference_id": "alibabacloud_ai_search_embeddings"
      }
    }
  }
}

Insert a lengthy Alibaba Group description and query it; the system returns relevant chunks, confirming automatic chunking works.