Integrate Anthropic Web Search Tool for Supported Models

[Dual-0]

Anthropic's Web Search tool enhances the capability of their models by providing access to current information, overcoming knowledge cutoffs. Implementing this feature in LibreChat would significantly improve the utility of supported Anthropic models within the platform, offering users more accurate, timely, and verifiable responses, especially for topics requiring recent information.

Technical Details (Based on Anthropic API Documentation):

API Endpoint: The web search functionality is provided via the standard messages API endpoint by including a tools parameter in the request payload.

Tool Definition: The tool is defined as follows:

{
  "type": "web_search_20250305",
  "name": "web_search",

  // Optional: Limit the number of searches per request
  "max_uses": 5,

  // Optional: Only include results from these domains
  "allowed_domains": ["example.com", "trusteddomain.org"],

  // Optional: Never include results from these domains
  "blocked_domains": ["untrustedsource.com"],

  // Optional: Localize search results
  "user_location": {
    "type": "approximate",
    "city": "San Francisco",
    "region": "California",
    "country": "US",
    "timezone": "America/Los_Angeles"
  }
}

LibreChat would need to allow users to enable this tool and potentially configure the optional parameters (max_uses, allowed_domains, blocked_domains, user_location).

API Flow:
1. User sends a message.
2. LibreChat constructs the API request including the tools array with the web search definition if enabled for the selected model.
3. The Anthropic API decides if a search is needed based on the prompt.
4. If a search is performed, the API returns server_tool_use and web_search_tool_result content blocks in the response.
5. LibreChat needs to parse these blocks to display the search activity (e.g., "Searching...") and the search results.
6. Anthropics final response will include text blocks with citations referencing the search results. LibreChat needs to parse and display these citations correctly.
7. Handle multi-turn conversations: Encrypted content/indices (encrypted_content, encrypted_index) from search results and citations must be passed back to the API in subsequent turns within the web_search_tool_result blocks to maintain context and citations.

Error Handling: LibreChat should handle potential web_search_tool_result_error responses and inform the user (e.g., max_uses_exceeded, too_many_requests, unavailable).

Streaming: If streaming is used, LibreChat must process streamed content_block_start and content_block_delta events for server_tool_use and web_search_tool_result blocks, including the pause while the search executes.

Usage/Pricing: Note that Anthropic charges separately for web searches ($10 per 1,000 searches) in addition to token usage. This might be relevant information for users.

Implementation Considerations:

• UI element to enable/disable web search for supported Anthropic models.
• Potential UI elements to configure optional parameters (max_uses, location).
• Parsing and rendering of server_tool_use, web_search_tool_result, and citations content blocks in the chat interface.
• Managing the passing of encrypted content/indices in multi-turn requests for persistent citations and context.
• Displaying search errors gracefully.

This feature would bring parity with the native Anthropic web search experience for users of supported models within LibreChat.

[Dual-0]

Closed by #8281