Skip to content

builtin-web-fetch.md

Latest commit

 

History

History
37 lines (28 loc) · 1.53 KB

builtin-web-fetch.md

File metadata and controls

37 lines (28 loc) · 1.53 KB
Raw

Built-in web fetch tool

WebFetchTool follows the veadk-python web_fetch contract. It performs a plain HTTP GET, extracts readable HTML or PDF text, and does not execute JavaScript.

var agent = Agent.builder()
        .name("reader")
        .model(new ArkLlm("your-model-endpoint"))
        .instruction("Use web_fetch to read public links before answering.")
        .tools(new WebFetchTool())
        .build();

The tool accepts these model-visible arguments:

  • url (required): public http or https URL.
  • extract_mode: markdown (default) or text.
  • max_chars: result limit, default 50000 and hard-capped at 200000.

Successful results contain url, title, content, and truncated. Failures contain error. No tool-specific credentials are required.

The implementation blocks private, loopback, link-local, multicast, and reserved targets after DNS resolution. HTTP and HTML meta-refresh redirects share a three-hop limit and every destination is validated again. Requests time out after 30 seconds; HTML/text downloads are capped at 2 MB and PDF downloads at 10 MB. Successful results are cached in process for 15 minutes.

The resolver validation and the HTTP connection are separate operations, matching the current Python baseline. Deployments that accept attacker-controlled URLs should additionally enforce egress policy at the network boundary to eliminate DNS-rebinding races.

For the AgentKit Java Runtime example, mount the tool by adding it to AGENTKIT_TOOLS:

export AGENTKIT_TOOLS=web_fetch