Uh oh!
There was an error while loading. Please reload this page.
docs: adapter protocol families + BYO + 3 platform-adapter integration guides (#302 Phase H) - #438
Conversation
…n guides (#302 Phase H) Adds the Phase H documentation subset, grounded in current ai-gateway origin/main code (every field name, secret shape, dispatch claim, and curl example verified against source): - reference/adapters.md — the 5 adapter protocol families (openai, anthropic, bedrock, vertex, azure-openai), the two-tier dispatch (specialized-by-`provider` then family-by-`adapter`), specialized bridges vs OpenAI-compat reuse, catalog vs BYO, featured vs non-featured, with a request-flow diagram. - configuration/byo-endpoint.md — bring-your-own OpenAI-compatible endpoint (vLLM / SGLang / Ollama / self-hosted proxy): provider + adapter:openai + api_base + secret, full curl flow, the Model `cost` block for budget tracking, and the standalone-doesn't-enforce-pricing caveat (cost_usd=0.0; cp-api recomputes server-side). - integration/upstream-bedrock.md — SigV4 JSON secret {access_key_id, secret_access_key, region, session_token?}; Anthropic `/invoke` (non-stream) vs Converse dispatch; streaming via Converse stream for all publishers; cross-region inference profile prefixes. - integration/upstream-vertex.md — service-account JSON → JWT → OAuth mint (and pre-minted access_token mode); Gemini `:generateContent`; api_base regional override; clearly-marked roadmap note that Anthropic/Llama-on-Vertex return a "not yet implemented" error. - integration/upstream-azure-openai.md — deployment URL + api_version; BOTH auth schemes (bare-string api-key vs JSON AAD {tenant_id, client_id, client_secret, authority_host?} → Bearer, the authority_host seam from ai-gateway#434); national-cloud authorities; content-filter tolerance; a curl example per scheme. Each page carries Docusaurus frontmatter (sidebar_position slots into its folder with no collision), a Verification section that exercises an observable contract (alias-restore on response.model + the outbound auth/URL shape) rather than just a 200, and the plaintext-secret production warning. Only the 5 new files are added; no existing pages or sidebar config touched.
ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Free Run ID: 📒 Files selected for processing (5)
📝 WalkthroughWalkthroughThis PR adds comprehensive documentation for AISIX AI Gateway's upstream adapter system and provider integrations. It establishes the architecture for routing requests to OpenAI-compatible endpoints and cloud providers (Azure, Bedrock, Vertex), with configuration guides, request/response mapping examples, and verification procedures for each integration pattern. ChangesUpstream Integration and Adapter Reference Documentation
Estimated code review effort🎯 2 (Simple) | ⏱️ ~10 minutes Note 🎁 Summarized by CodeRabbit FreeYour organization has reached its limit of developer seats under the Pro Plan. For new users, CodeRabbit will generate a high-level summary and a walkthrough for each pull request. For a comprehensive line-by-line review, please add seats to your subscription by visiting https://app.coderabbit.ai/login.If you believe this is a mistake and have available seats, please assign one to the pull request author through the subscription management page using the link above. Comment |
Audit (partial, before stall) flagged that adapters.md described the Vertex bridge as minting a token "before each call". That is imprecise on two counts, both already documented correctly in the upstream-vertex guide: (1) the `access_token` secret mode supplies a pre-minted token verbatim (no mint), and (2) the service-account mode caches the minted token (not per-call). Reworded the summary line to "minted in-process and cached, or supplied pre-minted" with a link to the Token minting section. No behavioral claim changed elsewhere.
moonming
commented
May 29, 2026
Independent doc-accuracy audit: clean — no HIGH, no MEDIUM. A fresh cold-read agent verified every factual claim across all 5 pages against
Two precision issues found during review were already fixed pre-merge: LOW (non-blocking, no change): the Converse-publisher list shows Merging on green CI. |
Summary
Phase H documentation subset for #302 — 5 new pages, each grounded in current
origin/maincode (every field name, secret shape, dispatch claim, and curl example verified against source):reference/adapters.mdprovider→ family-by-adapter) + specialized bridges vs OpenAI-compat reuse + catalog/BYO + featured/non-featuredconfiguration/byo-endpoint.mdprovider+adapter:openai+api_base+secret, Modelcostblock, standalone-pricing caveatintegration/upstream-bedrock.md/invokevs Converse dispatch, cross-region inference profilesintegration/upstream-vertex.mdintegration/upstream-azure-openai.mdclient_credentialswithauthority_host, the seam from #434), content-filter toleranceAccuracy verification (against origin/main)
Confirmed by reading source — not assumed:
Adapterenum ={Openai, Anthropic, Bedrock, Vertex, AzureOpenai}; dispatch viaresolve_bridge→dispatch_two_tier(crates/aisix-proxy/src/dispatch.rs).ProviderKeyhasprovider: String+adapter: Option<Adapter>;telemetry_tags.kind=catalog/byo.session_token?(bedrock/src/bridge.rs:268);chat_streamroutes all publishers (incl. Anthropic) through Converse stream; non-stream Anthropic uses/invoke.service_account_json/access_token; non-Gemini publishers return a "not yet implemented"Configerror (vertex/src/bridge.rs:533).AzureSecret::parsedetects api-key (bare string) vs AAD (JSON{);authority_hostoptional override (feat(azure-openai): production AAD authority_host override seam (closes #413) #434).cost_usd=0.0; cp-api recomputes server-side (chat.rs:908).Scope
Only the 5 new files. Each has frontmatter (no
sidebar_positioncollision), a Verification section that checks an observable contract (alias-restore + outbound auth/URL shape, not just 200), and the plaintext-secret production warning.Follow-ups (NOT in this PR — flagged for separate work)
reference/provider-compatibility.md+overview/feature-matrix.mdstill describe the old 6-value provider enum and don't mention the 5 adapter families — should backlinkadapters.md.configuration/provider-keys.mddocuments onlydisplay_name/secret/api_base—provider,adapter, andtelemetry_tagsare now on the struct; should be documented + backlink the 3 upstream guides.configuration/models.md(~line 160) has a stale claim thatprovideris a closed enum (openai/anthropic/google/deepseek/cohere/jina); on origin/main it's a free-form string — pre-existing doc bug to fix separately.🤖 Generated with Claude Code
Summary by CodeRabbit