文档首页

Support

Troubleshooting

Common failures—keys, connectors, widget CSP, empty retrieval—and how to narrow root cause before contacting support.

troubleshootingerrorsdebug

Most incidents cluster around credentials, scope, and stale content. This page offers a fast triage order; attach logs and timestamps when opening tickets.

Reproduce in the dashboard test chat to determine whether issues are channel-specific or agent-wide.

API and model key errors

401/403 often mean wrong key, wrong org, or expired BYOK credential. Verify environment (prod vs staging) matches the key you configured.

After rotation, restart or refresh dependent services that cache headers.

Poor or empty answers

Confirm the agent uses the knowledge base you expect. Re-run sync or crawl jobs if content recently changed.

Inspect whether instructions are overly restrictive—models may refuse even when retrieval hits.

Widget not loading

Check CSP, ad blockers, and script placement. Look for console errors referencing blocked connect-src hosts.

Confirm the embed key matches the agent and domain allowlists if enabled.

Upload or crawl quota errors (402 / “limit reached”)

If knowledge uploads fail with a quota or payment-required style error after many images or media files, you may have exhausted hosted monthly limits for Gemini image recognition or media transcription.

Mitigations: add a Google Gemini API key under Settings → LLM API Keys so those operations use your Google project, ask an administrator to increase plan limits, or temporarily upload text-native formats and disable media processing on crawls.

  • Crawl jobs that enable image or media processing require a resolvable Gemini key (hosted or org); otherwise the API rejects the crawl start with a missing-Gemini message.
  • OCR-only text from images may still appear when vision is skipped; do not assume the file failed entirely—check the stored document body.

在你的技术栈上构建

准备上线有依据的助手了吗?

开始试用,或与我们沟通上线、治理和企业级要求。