Lite edition

Product overview

The LiteAdvanced search engine is a high-performance, low-cost solution designed for Agent WebSearch scenarios. This engine builds high-quality, full-text web page indexes on demand and offers several core advantages:

• Low latency: Responds in under 500 ms on average, ensuring a smooth user experience for multi-turn Agent searches.

• Superior semantic search: Offers precise semantic search that outperforms traditional keyword-based search engines, achieving state-of-the-art (SOTA) performance on the ChineseSimpleQA dataset.

• Advanced search capabilities: Includes fine-grained filters for criteria like publish date and site.

• Low cost: Lowers your total cost of ownership.

Latency

The LiteAdvanced engine outperforms other common search engines, achieving an average response time (RT) of under 500 ms and a P95 latency of less than 600 ms. This performance gives it a competitive edge against general-purpose engines.

Quality

On a private dataset, the engine improved the overall_score by 13% compared to its predecessor, evaluated using a large model-based method for search quality.

The engine was also benchmarked on the public ChineseSimpleQA and SimpleQA datasets.

ChineseSimpleQA

Evaluation method

• Judge Model: qwen-max

• Answer Model: deepseek-v3

• Simplified logic for context enhancement with search results:

  def wrap_xml_context(search_results):
    """
        Wrap search results into XML format for context.
      """
    page_items = search_results.get("pageItems", []) if search_results else []
  
    if not page_items:
      return "<no_results>No relevant information found.</no_results>"
  
    xml_contexts = [
      f"<source>\n    <url>{item['link']}</url>\n    <passage>{item['snippet']}</passage>\n</source>"
      for item in page_items[:10]
    ]
      
    return "\n".join(xml_contexts)
  
  answer_prompt_template = """
      Please answer the following question based on the provided context.
      Question: {question}
      <context>
          {context}
      </context>
  """
  
  answer_prompt = answer_prompt_template.format(question = question, context = context)
  messages = [
    {"role": "system", "content": "You are an AI assistant. You use XML-formatted context to research people's questions."},
    {"role": "user", "content": answer_prompt}
  ]

Evaluation results

SimpleQA

Agent WebSearch requires high-quality, global, and multilingual content, distinguishing it from traditional search engines designed for humans. The SimpleQA benchmark is used to measure this global search capability. While many domestic engines lack sufficient multilingual support, LiteAdvanced is built on a high-quality global index.

Evaluation method

• Judge Model: qwen-max

• Answer Model: deepseek-v3

• The context enhancement logic is the same as that used for the ChineseSimpleQA test, but adds the --search-engine parameter.

• The test was conducted on the first 1,000 questions from the dataset.

 python -m simple-evals.simple_evals --model deepseek-v3 --eval simpleqa --examples 1000 --search-engine liteAdvanced

Evaluation results

Features

The engine supports several advanced search parameters and has been optimized for Agent-friendly operation.

Time range search

The engine supports time range search, letting you specify a start and end date for the publish date. This is in addition to the generic TimeRange parameter.

"advancedParams": {
  "startPublishedDate": "2024-12-01",
  "endPublishedDate": "2025-01-31"
}

Site search

To perform a site search, specify domains to include (includeSites) or exclude (excludeSites). Both top-level (e.g., sina.com.cn) and subdomains (e.g., finance.sina.com.cn) are supported, with a limit of 20 sites per request.

"advancedParams": {
  "excludeSites": "aliyun.com,sina.cn"
}

Number of results

You can specify the number of results to return, from 1 to 50, depending on your use case and result processing capacity. The default is 10.

"advancedParams": {
  "numResults": "10"
}

Semantic search

Unlike previous versions that required keyword rewriting, the new engine supports semantic search with natural language queries (ideally under 30 characters). For example:

• news about Zhejiang from the Two Sessions coverage

• positive news about NIO cars

Longer snippets

While traditional search snippets are short (~100 characters), Agent workflows benefit from richer context. To meet this need, the LiteAdvanced engine provides a longer snippet (up to 500 characters) to deliver more information in a single call. This reduces the need to choose between a short snippet, the full main_text (~3,000 characters), or a secondary web crawl.

Limitations

1. Performance for long-tail queries and real-time data needs (minute-level updates) may not match that of large-scale, traditional search engines. However, it outperforms other self-built search solutions.

Usage

For detailed instructions, see the IQS UnifiedSearch LiteAdvanced engine documentation. Set the engineType parameter to LiteAdvanced.

Release notes