 Hồ Xuân DũngandCursor
|
a57a8ce206
|
feat: add Gemini embeddings support + Letta compatibility fixes
Cherry-picked from decolua/9router#148 (author: xuandung38 / Hồ Xuân Dũng <me@hxd.vn>)
- Add Google AI (Gemini) embeddings support for /v1/embeddings endpoint
- Add Gemini embedding models: gemini-embedding-001, text-embedding-005, text-embedding-004
- Inject missing object/created fields for Letta and strict OpenAI clients
- Strip Azure-specific fields (prompt_filter_results, content_filter_results) from responses
- Fix Dockerfile: copy open-sse directory into Docker runner stage
Skipped: whitelist message field stripping (commit 3/7/8) — too aggressive for all providers
Skipped: default stream=false change (commit 9) — behavior change needs further review
Co-authored-by: Cursor <cursoragent@cursor.com>
|
2026-02-20 15:01:10 +07:00 |
|
|
HXD.VN
|
e1b836168a
|
feat: add /v1/embeddings endpoint (OpenAI-compatible) (#146)
* feat: implement /v1/embeddings endpoint (#117)
Add OpenAI-compatible POST /v1/embeddings endpoint that routes through
the existing provider credential + fallback infrastructure.
Changes:
- open-sse/handlers/embeddingsCore.js: core handler (handleEmbeddingsCore)
* Validates input (string or array), encoding_format
* Builds provider-specific URL and headers for openai, openrouter,
and openai-compatible providers
* Handles 401/403 token refresh via executor.refreshCredentials
* Returns normalized OpenAI-format response { object: 'list', data, model, usage }
- cloud/src/handlers/embeddings.js: cloud Worker handler (handleEmbeddings)
* Auth + machineId resolution identical to handleChat
* Provider credential fallback loop with rate-limit tracking
- cloud/src/index.js: wire new routes
* POST /v1/embeddings (new format — machineId from API key)
* POST /{machineId}/v1/embeddings (old format — machineId from URL)
* test: add unit tests for /v1/embeddings endpoint
- Setup vitest as test framework (tests/ directory)
- embeddingsCore.test.js (36 tests):
- buildEmbeddingsBody: single string, array, encoding_format, default float
- buildEmbeddingsUrl: openai, openrouter, openai-compatible-*, unsupported
- buildEmbeddingsHeaders: per-provider headers, accessToken fallback
- handleEmbeddingsCore: input validation, success path, provider errors,
network errors, invalid JSON, token refresh 401 handling
- embeddings.cloud.test.js (23 tests):
- CORS OPTIONS preflight
- Auth: missing/invalid/old-format/wrong key → 401/400
- Body validation: bad JSON, missing model, missing input, bad model → 400
- Happy path: single string, array, delegation, CORS header, machineId override
- Rate limiting: all-rate-limited → 429 + Retry-After, no credentials → 400
- Error propagation: non-fallback errors, 429 exhausts accounts
Total: 59/59 tests passing
Framework: vitest v4.0.18, Node v22.22.0
* feat: add Next.js API route for /v1/embeddings endpoint
Wire the embeddings handler into Next.js App Router.
- src/app/api/v1/embeddings/route.js: Next.js API route (POST + OPTIONS)
- src/sse/handlers/embeddings.js: SSE-layer handler mirroring chat.js pattern
Uses handleEmbeddingsCore from open-sse/handlers/embeddingsCore.js with
the same auth, credential fallback, and token refresh logic as the chat
handler. Supports REQUIRE_API_KEY env var, provider fallback loop, and
consistent logging.
|
2026-02-18 13:24:02 +07:00 |
|