[fahreddinozcan]

Documents the on-prem decision model setting from upstash/context7parser#1058 (CTX7-3710).

• New Features page enterprise/decision-models: where a decision model replaces the language model (page classification, duplicate checks, reranking) and the fallback.
• Steps to turn it on, the settings, the four providers (OpenRouter, OpenAI, Vercel AI Gateway, Custom System One endpoint) with default models, and the test button.
• Note on thresholds tuned for Jev 1.13, and a warning on what data goes to the provider.
• Getting Started > Configuration gets a short "Decision Model" pointer to the new page.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
context7-admin-mcp	Ignored		Nov 10, 2026 9:29am UTC

[linear-code]

CTX7-3710

[context7]

Docs7 for context7.com

Result	Status	Action
Deployment	✅ Ready	Open preview
Content review	✅ Passed. No problems found.	—

Commit 45f1c98 · Updated 2026-11-10 09:30 UTC · View build details

[ramazanoacar]

Thanks Fahreddin, clear and well scoped. The table of which decisions move and which stay with the language model, the fallback line, and the data warning answer the main questions an admin would have. The Jev 1.13 note is a good catch too, since OpenAI is the odd one out. Approving 👍

Small notes (non-blocking)

• Version: the page doesn't say which image release adds the setting, and the changelog doesn't mention it yet. A short "Available from vX.Y" (or a changelog entry when it ships) helps admins on 2.0.0 who don't see the toggle.
• Network: on-prem customers often run with an egress allowlist. Listing the host each provider calls (OpenRouter, OpenAI, AI Gateway) next to the Providers table would save them a support ticket, like the Network access table on the Google page.
• Reranking now calls the provider on each search. If there's a timeout before it falls back to the language model, mentioning it (or that it adds a network round trip) would set expectations on latency.
• Nits: "answers yes/no questions" doesn't quite fit reranking or "what each page contains"; maybe "makes short classification and ranking decisions". And "Stored encrypted in the workspace database" — the rest of the docs say "application database".

[enesgules]

Approve. The page matches #1058: default models (typesafe/jev-1.13-20260917, gpt-6-luna, typesafe-ai/jev-1.13.0, jev-1.13), the Use a decision model switch, Test decision model (one duplicate check), the fallback to the language model, and the rerank-only-when-on rule. CI is green.

Nit: after #1066 the nav item is Workspace Settings (it shows "Settings" only on narrow screens), so step 1 could say Workspace Settings > AI models.