Docs: troubleshoot AI responses truncated by infrastructure timeouts
What does this MR do?
Documents a failure mode that has no coverage today: a long AI response, particularly from a large or reasoning model, cut off partway through, often with no error shown. This happens when a proxy, load balancer, or firewall between components enforces a request or idle timeout shorter than the response and severs the stream mid-flight.
Six small additions, one topic:
- A new symptom entry in GitLab Duo Self-Hosted troubleshooting, "Responses are truncated without an error", walking the timeout at each hop. Complements gitlab-org/modelops/applied-ml/code-suggestions/ai-assist!6086 (merged), which made the chat model request timeout configurable.
- A streaming-timeout bullet in the general Duo outbound-connections section (applies to any self-managed instance behind an egress proxy, GitLab-hosted AI Gateway included).
- Comments on
proxy_read_timeoutin both NGINX reference configs ininstall_ai_gateway.md, clarifying it bounds the idle gap between response chunks, not total response time. - One added pointer at the end of the Error A1000 resolution list (truncation without an error is the sibling symptom of A1000).
- One cross-link from "Determine the timeout value".
- One breadcrumb from Error A1003 in GitLab Duo Chat troubleshooting to the new material.
Related
- gitlab-org/modelops/applied-ml/code-suggestions/ai-assist#2216 (silent truncation at the gateway layer)
- gitlab-org/modelops/applied-ml/code-suggestions/ai-assist!5833 and gitlab-org/modelops/applied-ml/code-suggestions/ai-assist!6029 (platform timeout raise, in flight)
Author checklist
- Optional. Consider taking the GitLab Technical Writing Fundamentals course.
- Follow the:
- If you're adding or changing the main heading of the page (H1), ensure that the product tier badge is added.
- If you are a GitLab team member, request a review based on:
- The documentation page's metadata.
- The associated Technical Writer.