feat: cloud provider catalog (Azure OpenAI, Bedrock, Vertex) + embeddings capability (§3.4)#335
Open
gnanirahulnutakki wants to merge 6 commits into
Open
feat: cloud provider catalog (Azure OpenAI, Bedrock, Vertex) + embeddings capability (§3.4)#335gnanirahulnutakki wants to merge 6 commits into
gnanirahulnutakki wants to merge 6 commits into
Conversation
added 2 commits
July 14, 2026 02:09
Adds the text-embedding sibling of the §3.0 Provider trait: a separate object-safe EmbeddingProvider (embed/id/embedding_dim/rate_card) plus an EmbeddingProviderRegistry, matching ProviderRegistry's shape. Kept separate from Provider rather than added as a method, since not every backend offers both capabilities (Anthropic has no embeddings endpoint). Checkpoint: architect/sessions/cloud-providers-embeddings-2026-07-14/journal.md Signed-off-by: Gnani Nutakki <gnani.nutakki@gmail.com>
No prior blueprint covered these three clouds — plans/3.0-3.3 and 12.5 cover Anthropic/OpenAI/Ollama/OpenRouter/Claude-CLI only, and Bedrock/Vertex/Azure appeared solely as an unwritten future-catalog name-check. Authored plans/3.4 (local design doc; plans/ is gitignored here and synced by the docs-resync lane) and built these three Provider implementations against it: - ardur-provider-azure-openai: real HTTP against Azure's resource/deployment -scoped Chat Completions + Embeddings endpoints (api-key header). - ardur-provider-bedrock: real HTTP against Bedrock Runtime's InvokeModel for Anthropic Claude models, authenticated with a hand-rolled AWS SigV4 signer (sha2+hmac, already workspace deps — no aws-sdk/aws-sigv4 pulled in for one signed POST). - ardur-provider-vertex: real HTTP against Vertex's generateContent for Gemini models. Standard Vertex auth needs an RS256 service-account JWT and the workspace has no RSA crate; Phase 1 takes a pre-minted bearer token via ARDUR_VERTEX_ACCESS_TOKEN instead of adding new asymmetric-crypto deps without separate review. All three implement tool-calling via the shared ToolDef/ToolCall envelope; supports_streaming() is false for all three (Phase 2 TODO — each cloud's incremental path is a distinct wire protocol from the shared SSE type), so Provider::stream's trait default wraps one complete() call, matching the non-streaming precedent provider-codex/provider-claude-cli already set. Wires all three into provider-selector (ARDUR_PROVIDER=azure-openai|bedrock |vertex). Each crate ships unit tests plus a wiremock e2e-stub test of its public surface; provider-selector gains matching from_env integration tests. cargo build --workspace / clippy --workspace --all-targets -D warnings / test --workspace all green. Checkpoint: architect/sessions/cloud-providers-embeddings-2026-07-14/journal.md Signed-off-by: Gnani Nutakki <gnani.nutakki@gmail.com>
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
…ex (§3.4) Closes the streaming-parity gap left by the initial cloud-provider drop: all three previously wrapped Provider::stream's trait default (a single-shot complete() call). Now each has a real incremental decoder and supports_streaming() == true: - Azure OpenAI: hand-rolled line-buffered SSE decoder (wire framing is byte-identical to OpenAI's) — no eventsource-stream dependency, matching provider-openai-compat's established pattern. - Vertex: streamGenerateContent?alt=sse (Gemini's documented SSE mode), same line-buffered approach. Gemini streams whole functionCalls per chunk rather than incremental argument fragments. - Bedrock: InvokeModelWithResponseStream answers with AWS's binary application/vnd.amazon.eventstream framing, not SSE — a new eventstream.rs hand-rolls the generic frame decoder (length-prefixed messages + CRC32, checked against the standard "123456789" test vector; no crc32fast/ aws-eventstream crate pulled in), and streaming.rs maps the base64-wrapped Anthropic-native streaming events (message_start/content_block_start/ content_block_delta/content_block_stop/message_delta/message_stop) onto the shared StreamEvent protocol. Every provider ships a full E2E streaming test driving the real public Provider::stream() method through actual wiremock-served SSE/event-stream bytes (Bedrock's through a real SigV4-signed, CRC-valid binary frame feed) — not just unit tests of the decode functions in isolation. cargo build --workspace / clippy --workspace --all-targets -D warnings / test --workspace all green (317 test-result blocks, 0 failures). Checkpoint: architect/sessions/cloud-providers-embeddings-2026-07-14/journal.md Signed-off-by: Gnani Nutakki <gnani.nutakki@gmail.com>
…ry and parse_retry_after_ms Signed-off-by: ArdurAI <team@ardur.ai>
…mbeddings-2026-07-14 Signed-off-by: ArdurAI <team@ardur.ai>
Signed-off-by: ArdurAI <team@ardur.ai>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Extends ardur-agent's model/LLM provider breadth per a lane assignment to own this domain.
plans/3.0already unifies theProvidertrait andplans/3.1-3.3/12.5already implement Anthropic, OpenAI-compatible, Ollama, OpenRouter, and Claude Code CLI — all left untouched. No blueprint existed for AWS Bedrock, Google Vertex, Azure OpenAI, or an embeddings capability (they appeared only as an unwritten future-catalog name-check), so aplans/3.4design doc was authored first (local-only;plans/is gitignored in this repo and synced by the docs-resync lane) and these crates were built against it:ardur-provider-runtime: newEmbeddingProvidertrait (separate fromProvider— not every backend offers both capabilities) +EmbeddingProviderRegistry.ardur-provider-azure-openai: real HTTP against Azure's resource/deployment-scoped Chat Completions + Embeddings endpoints, plus real incremental SSE streaming (a hand-rolled line-buffered decoder — Azure's streaming wire framing is byte-identical to OpenAI's).ardur-provider-bedrock: real HTTP against Bedrock Runtime'sInvokeModel/InvokeModelWithResponseStreamfor Anthropic Claude models, authenticated with a hand-rolled AWS SigV4 signer (built onsha2+hmac, already workspace deps — noaws-sdk/aws-sigv4pulled in). Streaming decodes AWS's binaryapplication/vnd.amazon.eventstreamframing (length-prefixed frames + CRC32 — hand-rolled, checked against the standard test vector) and maps the base64-wrapped Anthropic-native streaming events onto the shared protocol.ardur-provider-vertex: real HTTP against Vertex'sgenerateContent/streamGenerateContent?alt=ssefor Gemini models. Standard Vertex auth needs an RS256 service-account JWT and the workspace has no RSA crate, so Phase 1 takes a pre-minted bearer token viaARDUR_VERTEX_ACCESS_TOKENrather than adding new asymmetric-crypto dependencies without separate review (flagged as a Phase 2 TODO).ardur-provider-selector: wired all three in (ARDUR_PROVIDER=azure-openai|bedrock|vertex).All three new providers implement tool-calling via the shared
ToolDef/ToolCallenvelope and now have real incremental streaming parity with the existing five providers (supports_streaming() == true, no longer wrapping the trait-default single-shotcomplete()call).Test plan
cargo test -p ardur-provider-runtime— EmbeddingProvider traitcargo test -p ardur-provider-azure-openai— 14 unit + 4 wiremock e2e (chat + embeddings + streaming + 401 mapping)cargo test -p ardur-provider-bedrock— 21 unit (incl. SigV4 signer + event-stream frame decoder + CRC32) + 3 wiremock e2e (complete + tool_use + full sign→frame→decode streaming round-trip)cargo test -p ardur-provider-vertex— 12 unit + 4 wiremock e2e (complete + function-call + streaming + 401 mapping)cargo test -p ardur-provider-selector— 12 unit + 12 env-integration (incl. 3 new azure/bedrock/vertex cases)cargo fmt --check— cleancargo build --workspace— greencargo clippy --workspace --all-targets -- -D warnings— greencargo test --workspace— green (317test result: okblocks, 0 failures, verified via direct exit code not a piped tail)No live credentials required for any of the above — every provider ships a wiremock-backed stub test (including real SSE / binary-eventstream bytes for the streaming paths), matching the §3.1 pattern. Live-credential smoke tests, Bedrock/Vertex embeddings, and Vertex local service-account JWT minting are documented Phase 2 follow-ups.