[marcus-souza-azion]

What & why

Related issue: MM-16308 (English) and MM-16309 (Portuguese), in Jira. Their descriptions list every change by page.
Pages affected: 64 pages (32 en, 32 pt-br) and src/nav/trees/ai-inference.json. One page pair is new. No existing permalink changes.

• AI Inference, under /en/documentation/platform/ai-inference/, in both languages: / (the overview), quickstart/, how-it-works/, model-invocation/, models/, limits/, best-practices/, lora-fine-tune/, glossary/, guides/, the new troubleshooting/ (/pt-br/documentacao/plataforma/ai-inference/solucao-de-problemas/), and the model pages qwen3-30ba3b/, gpt-oss-20b/, mistral-3-small/, qwen-2-5-vl-3b/, qwen-2-5-vl-7b/, nanonets-ocr-s/, qwen3-embedding-4b/, baai-bge-reranker-v2-m3/
• The two owner pages of the section's mechanism, in both languages: /en/documentation/use-cases/build-and-run-ai-workloads/ai-inference-architecture/ and /en/documentation/devtools/runtime/api-reference/ai/
• Pages outside the section that contradicted the verified facts, in both languages:
  • Cells, in six pages: /en/documentation/devtools/runtime/how-it-works/, the Azion Runtime glossary, the Functions glossary, Azion Runtime debugging, /en/documentation/devtools/runtime/node/, and Orchestrator Edge Services
  • /en/documentation/guides/application-development/frameworks/ai-inference-starter-kit/
  • the gateway guide route-model-calls-through-a-gateway-function/ and the use case govern-access-to-multiple-ai-models/
  • the guide embed-documents-into-a-vector-table/
  • About Azion Console, getting-to-know-azion-console/

Part 1: live tests in an account with access to AI Inference

The section had no live evidence: an earlier probe could not tell a wrong id from an account without access. This run used an account with access, a probe Function that passed the id and body to Azion.AI.run and returned the result unchanged, and one deployment of the AI Inference Starter Kit template from Azion Console. It settled 52 facts, and the pages now state them:

• Access. On the Hobby and Pro plans, AI Inference is available upon request, subject to approval; on Enterprise, it is custom. Source: the pricing page.

• The binding.

  • Azion.AI.run takes the model id first and the body second, and no credential.
  • A model field in the body is ignored, and so is a field the model does not know.
  • With stream: true, the call resolves to an async iterator of chat.completion.chunk objects, with no [DONE] marker.
  • Azion.AI.models() lists the ids the account can call.
  • A wrong id and an account without access both fail with AIError: model not found or not allowed.

• Model ids. Each model answers to two forms, a repository path and a lowercase form. The model field of a response carries the lowercase form, so it can differ from the id sent. gpt-oss-20b is refused: GPT-OSS 20B answers to openai/gpt-oss-20b.

• Models.

  • The two Qwen2.5 VL models call a tool only when tool_choice names it. Leaving the choice to the model fails with 400.
  • Nanonets-OCR-s takes no tools.
  • Qwen3 Embedding 4B accepts dimensions 4096 but returns 2560 values.
  • The reranker refuses the text_1/text_2 scoring pair.
  • An image URL that serves no image fails with 500.

• The template. It creates:

  • a Workload with its deployment;
  • an Application with Cache, Functions, and Application Accelerator on;
  • a Function with its Function Instance;
  • one rule.

  Its Function reads the model from the model field of the body, answers on any path, takes only POST with Content-Type exactly application/json, and has no authentication in front. Application Name refuses a hyphen. Until the deployment propagates, the domain answers 404.

• Limits. The wait inside Azion.AI.run does not count against the Function's 2-second CPU limit.

Part 2: the section reviewed with the writing rules

Every page went through the full review of the documentation writing rules, and the review's findings were applied. Phase 1's facts are the source of truth. A claim that no test or Azion source supports was narrowed or removed.

• Cells. "Azion Cells" is an internal name, and the pages used it as a layer between the Function and the model. It is gone: the chain is now Function, then Azion.AI.run, then AI Inference. "Azion Runtime" appears only where a sentence says where the Function runs.
• Definitions. AI Inference, Azion.AI.run, model id, the OpenAI-compatible API, and LoRA Fine-Tune now read the same on every page that defines them. AI Inference is "a Platform Resource that runs open-source AI models on Azion's distributed infrastructure, in response to the calls your Functions send".
• Definitions owned by other sections. Workloads, deployment, workload domain, custom domain, and Application follow their owner pages. Workloads and Fundamentals are still in review, so these were checked against their review branches, not main:
  • the deployment "sets" the Application;
  • a custom domain answers in addition to the workload domain;
  • the domain belongs to the Workload, not the Application.
• Rewrites. The Overview, Glossary, and Best practices failed the structural review and were rewritten in full. Best practices is now "Best practices for AI Inference", one section per practice, each with a specimen. The other pages were corrected finding by finding.
• Diagrams. The Overview and How it works diagrams follow the section convention: a legend, numbered arrows, dotted responses, and the Workload as the entry point.
• Model pages. Licenses, sizes, and descriptions now come from each model's Hugging Face card. Example responses are the observed ones, not illustrations.
• Sidebar. Pricing and Changelog leave the Management group and sit at the top level after Glossary.

Part 3: the pages outside the section

These pages carried the same wrong facts, so they are corrected here instead of in a follow-up PR.

• Cells. Six Runtime, Functions, and Orchestrator pages now say "an isolation environment based on V8 isolates". The Functions glossary loses its Cell row. Literal values stay: azion logs cells, cellsConsoleEvents, and ext:cells_fetch.
• AI Inference Starter Kit guide.
  • The code example with a fixed model id is replaced by what the installed code does, read from the code itself.
  • The response's model is the lowercase id.
  • The domain answers 404 before it propagates; the guide used to say it "does not resolve".
  • The steps name Start from Template and Application Name, with no hyphen.
  • The menu paths name only the Console menu item, Functions and Applications, per the click-path rule of the style guide. The Quickstart follows the same rule.
• Model ids and widths.
  • The gateway guide and the governance use case use openai/gpt-oss-20b.
  • The embeddings guide says that 4096 returns 2560 values.
  • The customer-support use case was already correct and is unchanged.
• About Azion Console. The four-tab "Create modal" is replaced by Creation Center, with its Import from GitHub and Start from Template sections, read from the Console's public code and confirmed in the Console. The Console menu table comes from main (docs(style-guide): standardize Azion Console click paths #2596), and it matches the menu in that code.

Part 4: a Troubleshooting page

The section had no Troubleshooting page. Troubleshoot AI Inference is new, written from the symptoms the live tests reproduced, with one entry per symptom: the symptom, its cause, the fix, and the outcome.

• Binding errors:
  • model not found or not allowed, including a model value reused from a response;
  • a validation error;
  • Request failed with status 400 and 500;
  • GPT-OSS 20B with no content;
  • an embedding with 2560 values;
  • a field with no effect;
  • Azion.AI undefined under azion dev.
• AI Inference Starter Kit:
  • the endpoint's Request must be a POST request with JSON body, Missing 'model' field, Invalid JSON, and AI internal error;
  • the 404 right after the deployment;
  • Application Name refusing a hyphen.
• What it links instead of repeating: the full error bodies stay in Model invocation, Errors, and each entry links them.
• What stays out: symptoms no test reproduced, such as a terminated or deprovisioned model.
• Sidebar: the page sits between Best practices and Glossary, as in the Workloads sidebar.

URL and anchor changes

• No existing permalink changes, and no page moves.
• New page: /en/documentation/platform/ai-inference/troubleshooting/ and /pt-br/documentacao/plataforma/ai-inference/solucao-de-problemas/.
• Titles: Best practices is now "Best practices for AI Inference" ("Boas práticas para AI Inference").
• Headings whose anchor changed:
  • Overview: Invocation path → Request path; Model call → Function code.
  • Quickstart: Open the application domain → Open the workload domain; (Optional) Change the model the function calls → (Optional) Call another model.
  • Best practices: every section heading, for the rewrite.
  • BAAI reranker: Scoring is removed.
  • About Azion Console: Create modal → Creation Center.
  • New headings, such as Model invocation Errors, Limits Access by plan, and Runtime API List models.
• The anchor check covers the whole corpus: 4,299 anchor links resolve and none breaks, including the links into the changed pages and the links of the new page.

For review

• InternVL3 keeps a refused id. The binding refuses opengvlab-internvl3-1b-instruct, and no other id answers. The page was left unchanged on purpose.
• Best practices, practice 1, is half verified. With no INFERENCE_SECRET variable, a request with no header and one with Bearer undefined both receive 401, as the page says. The path with the correct secret needs that variable in a test account and was not run.
• Two owner pages disagree.
  • The Applications overview says a rule runs the Function.
  • Function Instances says the rule invokes the Function Instance.
  • The AI Inference pages follow the Applications overview.
• The Architecture page has no catalog entry. The catalog of Reference Architectures has no entry for it.
• The Runtime API page. Its facts are corrected, but its structure does not follow the API reference shape of the writing rules.

Outside this PR, unchanged

• The Functions overview still says AI Inference runs the model on Azion Runtime, that an OpenAI client "works unchanged", and that Run Function requires Application Accelerator.
• The Build AI agents use case lists only the system, user, and assistant roles, and leaves out tool.
• The Use the + Create button guide still describes four tabs.
• The Portuguese pricing page carries two ### AI Inference sections.
• Release notes and one sample header value still say Cells; they are records and literals.
• The pages fixed in Part 3 keep their older findings outside the changed lines: two container headings and a skipped heading level in Edge Services, a 324-character Portuguese Edge Services description, and the Portuguese debugging title.

How it was checked

• Write-time gate: no finding on a changed line, and none on the new page pair. The six pre-existing findings above are on lines this PR does not touch.
• Link check: no broken link and no cross-language link in the 64 pages, and every namespace is paired.
• Anchor check: 4,299 anchor links across the corpus, none broken.
• pnpm lint:navcheck: passes.
• ESLint and Prettier: ESLint is clean on the section pages, and Prettier is clean on the Part 3 pages.
• EN and PT parity: headings, fences, steps, and <Code> blocks match in every section pair. The Part 3 pairs got the same edits sentence by sentence.
• Dev server: the 18 section pages were opened in both languages, and the three diagrams render with no parse error. The Starter Kit guide, About Azion Console, Node.js APIs, and the Functions glossary were opened after Part 3. The Troubleshooting pair renders all 14 sections and appears in the sidebar, and its .md versions carry the same 14 sections with no MDX left in them.
• Live:
  • Runs on November 11 and 12, 2026, with Azion CLI 4.24.0, API v4, and Azion Console.
  • Every object created for the tests was deleted, and the account inventory matches the one before.
• Not verified:
  • pnpm build:local.
  • The .md Markdown versions of the pages, except the Troubleshooting pair.
  • How to create a LoRA adapter: no public interface creates one.

Type of change

• 🆕 New content (feat)
• 🩹 Fix (fix) — typo, broken link, wrong information
• ♻️ Content update (docs) — rewrite, expansion, upkeep
• 🌐 Translation sync (i18n)
• 🏗️ Platform / structure (refactor / chore) — reviewed by UXE, no content mixed in

Author checklist

• PR title follows type(scope): summary (see GOVERNANCE.md §4)
• Frontmatter complete: title, description, meta_tags, namespace, permalink, last_reviewed — every file has the first five fields, but none has last_reviewed, which no page in the corpus has. This PR adds one page pair with a new namespace and permalinks, and changes no existing permalink or namespace
• No legacy "edge-" product names in the copy — the only ones in changed lines are Edge Nodes and Edge Services, the Orchestrator objects as Azion Console names them
• How-to/tutorial content includes at least one runnable, copy-paste-tested code block — the Quickstart curl request ran as written against the template's endpoint on November 11, and the Best practices practice 1 Function ran as written for its refusal path
• Screenshots (if any) have alt text and follow image standards — no screenshot is added or changed
• Internal links are relative and resolve locally — internal links are root-relative with the language prefix, and the link check finds no broken link from the changed pages
• If any permalink changed or page moved: redirect added in this PR — no permalink changed and no page moved
• i18n: pt-br updated in this PR or follow-up i18n issue created: — every en change has its pt-br version in the same commit
• I ran pnpm build:local (build + frontmatter check) without errors