Skip to main content
This guide details how to leverage Web Search parameters effectively to optimize performance, relevance, and token usage for LLM-integrated applications.

​
Key Benefits

  • Semantic Intelligence: Hybrid keyword and semantic search ensures high relevance from natural language queries without complex engineering.
  • Context Control: Granular filters (domain, time, content) reduce noise and align results with specific use cases.
  • Cost & Latency Optimization: Retrieving highlights instead of full content minimizes token consumption, essential for Agentic workflows.

​
Search Configuration

  • Safety (safesearch):
    • "strict" (Default): Excludes adult content. Standard for public-facing chatbots to prevent toxic content generation.
    • "off": Unfiltered results. Required only for specialized research agents (e.g., medical or biological studies) where standard safety filters might incorrectly flag necessary anatomical or scientific content.

​
Precision Filtering

  • Domain Filtering (include_domains / exclude_domains): Functions as a whitelist or blacklist.
    • Trusted Knowledge Base: Using include_domains to restrict retrieval to high-authority sources (e.g., official documentation, .gov, or .edu sites) creates a “walled garden” that significantly reduces hallucination risks in professional contexts.
    • Noise Reduction: Using exclude_domains to filter out user-generated content platforms or content farms prevents the LLM from ingesting colloquial or unverified information.
  • Content Constraints (include_text / exclude_text): Enforces or forbids specific keywords within the page content. For example, requiring “quarterly earnings” to appear when searching for financial reports, or excluding “rumor” to filter out speculative content.
  • Time Sensitivity:
    • Breaking News Mode: Combining time_basis: "published" with a strict start_time (e.g., past 24 hours) forces the engine to ignore SEO-optimized evergreen content. This strategy is essential for news summarization or market analysis agents.
  • Result Count (count): Defaults to 5. For direct Q&A tasks, retrieving 3-5 results typically offers the best balance between context availability and latency. Higher counts (10+) are recommended for broad topic aggregation tasks.

​
Response Content & Format

  • Highlights vs. Full Content:
    • Highlights (Default): Returns relevant, concise snippets. This is the most token-efficient format for Fact-Checking and Q&A, where the answer is likely contained in a single paragraph.
    • Full Content: Returns parsed page text. Necessary for “Reading Assistant” agents that need to summarize entire articles, analyze writing style, or extract scattered data points from a long report.
    • Hybrid Strategy (Highlight-First): A cost-effective pattern involves requesting highlights first to assess relevance, and then triggering a second request for full_content only on the specific high-value URLs.
  • Output Format (format): Controls the output format of the highlight snippets.
    • "text" (Default): Returns plain text.
    • "markdown": Returns with basic formatting (e.g., bolding of matching terms) where supported.

​
Performance and Usage Considerations

  • Token Economy: The meta.usage field monitors consumption. To minimize operational costs, applications should default to highlights and only request full_content when user intent explicitly requires deep reading.
  • Metadata Utilization: Fields like time_published should be used for secondary ranking on the client side (e.g., prioritizing the absolute newest article among the top 5).