[danehans]

Company

Not applicable.

Contribution summary

Add standalone vLLM Semantic Router integration pages, reusing the shared Kubernetes overview to avoid drift.

Correct the previous claim that vSR output is available to HTTPRoute matching: matching occurs before the vSR processing used by the single-runtime example, while model routing consumes the selected model afterward.

Gate the v1.5.0 single-runtime example and its related guidance out of older documentation versions, and retain Kubernetes-specific reference links.

This PR depends on agentgateway/agentgateway#3471.

[cloudflare-workers-and-pages]

Deploying agentproxy with    Cloudflare Pages

Latest commit:	0c2865c
Status:	⚡️  Build in progress...

View logs

[danehans]

@artberger @kristin-kronstain-brown @Nadine2016 PTAL

[kristin-kronstain-brown]

Reusing the shared overview instead of forking a standalone copy is the right call, and the new trusted-tier-context bullet is a useful security note.

Blocking

• This shared page is also reused by kubernetes/1.3.x and 1.4.x. The single-runtime card and its AgentgatewayModel sentence now render there, but the example README requires v1.5.0. The PR also removes the previous "1.3.0 or later" note. Can you wrap that card, its sentence, and the "Trusted tier context" bullet in {{< conditional-text exclude-if="1.0.x,1.1.x,1.2.x,1.3.x,1.4.x" >}}? Also, "the matching agentgateway version" would be clearer as "agentgateway v1.5.0 or later".

Those versions still get pulled into enterprise docs, so it's good to keep them accurate even if they aren't in main or latest for oss.

Should fix

• The model aliases, request-body validation, and OTel stack links were removed, which fixes standalone (those pages do not exist there), but they still exist for Kubernetes. Can you keep them for Kubernetes with include-if="kubernetes"?
• The old page said PreRouting makes the vSR result available to HTTPRoute matching. The new k8s single-runtime README says HTTPRoute matching runs before vSR and cannot use its output. If the README is correct, this PR fixes an error, which is worth saying in the PR description. Either way, can step 1 still name the PreRouting phase for Kubernetes readers?
• On standalone, "Choose the one that matches the models and outcome" sits above a single card. Can you gate that sentence per product? Or revise that description?
• The prose now says "response caching", but the card title still says "Configure semantic caching".
• The standalone "For inference with vLLM, see Custom providers" line points to a page that does not mention vLLM. One clause about vLLM's OpenAI-compatible API would connect the two.
• Readers only see one deployment mode, so "The following flow applies to both standalone and Kubernetes deployments" can go.

[danehans]

@kristin-kronstain-brown thanks for the review. I updated the PR based on your feedback, PTAL.