[RaviTharuma]

Summary

• One Jina credential pool: dashboard jina-ai (preferred) shares a token with jina-reader and jina-search; JINA_AI_API_KEY / JINA_API_KEY are a real fallback only when no dashboard key exists (connection_id=env:JINA_AI_API_KEY).
• Jina v5 Omni multimodal is first-class: native EmbeddingsV5Request docs ({ text }, { image: url|base64 }, { content: [...] }, plus task / normalized) pass through POST /v1/embeddings intact to https://api.jina.ai/v1/embeddings. Catalog ids stay jina-ai/…; jina-embeddings-v5-omni aliases to jina-embeddings-v5-omni-small.
• Gemini Embedding 2 multimodal (GA April 2026, not “Gemini 5”): live catalog is gemini/gemini-embedding-2 (alias google/gemini-embedding-2). Dashboard gemini keys are used first; GEMINI_API_KEY / GOOGLE_API_KEY only when none exist. Native parts (text / inline_data / file_data / fused parts) are forwarded. N OpenAI input items become N :batchEmbedContents vectors — one fused Content still correctly yields one vector. Default dim 3072.
• UI labels distinguish Foundation API (api.jina.ai) from Reader (r.jina.ai). Dashboard Test probes GET /v1/models and falls back to jina-embeddings-v5-omni-small. Embed/rerank logs keep connection_id.
• New Jina proxies that fit OmniRoute: POST /v1/classify, POST /v1/segment, jina-search (s.jina.ai) with jina-ai / jina request aliases. Catalog adds jina-reranker-v3.5.

Fixes / implements #10579.

Related Issues

• Closes [BUG] Jina complete + Jina Omni multimodal + Gemini Embedding 2 multimodal via OmniRoute #10579
• Related to fix(api): accept Jina native {image: data:url} embedding input #10567 (native image normalize — this PR instead passes through Jina-native docs)
• Related to fix(api): name working OpenRouter ids when Gemini embed creds are missing #10565 (Gemini embed creds hint — this PR actually uses dashboard/env Gemini embed keys)

Validation

• Change type: provider
• Focused tests: jina-omni-multimodal, jina-complete-provider, gemini-embedding-2-multimodal, embeddings multimodal, search registry/route, provider-models, sse-auth Jina pool
• npm run check:env-doc-sync
• Reconciled with release/v3.8.50
• Production-code changes include automated tests

Tests Added Or Updated

• tests/unit/jina-omni-multimodal.test.ts — native text+image URL forwarded intact; no image fetch; family alias; text-only SKU rejects images
• tests/unit/jina-complete-provider.test.ts
• tests/unit/gemini-embedding-2-multimodal.test.ts — catalog alias; native parts forwarded (not string[]); N strings → N batch vectors; fused Content → one vector; 001 stays on OpenAI shim; env fallback helper
• tests/unit/embeddings-multimodal-7956.test.ts — Gemini N canonical items → batchEmbedContents
• tests/unit/embedding-family-guard.test.ts — embedding-2 default 3072

Coverage Notes

• Jina native pass-through covered by jina-omni-multimodal.test.ts.
• Gemini native transport + N→N covered by gemini-embedding-2-multimodal.test.ts.
• Auth pool + env fallback covered by the Jina/Gemini provider helper tests and sse-auth.test.ts.

Reviewer Notes

• Reader stays a separate card because the host is r.jina.ai. Labels make that obvious; the same token is reused.
• Jina not fitted (wrong OmniRoute model): chat/VLM, DeepSearch, batch embeddings, classifier train/CRUD.
• Jina-native { image: "https://…" } is forwarded after a public-HTTPS check; Jina fetches the URL. Canonical { type, source: { type: "url" } } still inlines server-side as before.
• Gemini Embedding 2 uses x-goog-api-key against :embedContent / :batchEmbedContents. Multiple parts inside one Content still fuse (Google’s documented aggregation). Multiple OpenAI input array elements do not fuse.
• google/gemini-embedding-2 is an exact model-string alias to gemini/gemini-embedding-2. A custom provider_node whose prefix is google plus gemini-embedding-001 is left unaliased.
• Changelog fragment: changelog.d/features/10581-jina-complete-provider.md.

Test plan

• POST /v1/embeddings with jina-ai/jina-embeddings-v5-omni-small and input: [{ "text": "…" }, { "image": "https://…" }] reaches api.jina.ai with those objects unchanged
• POST /v1/embeddings with gemini/gemini-embedding-2 (or google/gemini-embedding-2) and input: ["a", "b", "c"] returns 3 vectors via :batchEmbedContents
• Gemini native { text } + { inline_data } batch is forwarded (not stripped to strings); fused { parts: [text, image] } returns 1 vector
• Dashboard gemini key is used when present; GEMINI_API_KEY only when no dashboard row exists
• Bare / family ids: jina-embeddings-v5-omni, jina-embeddings-v5-omni-nano, jina-reranker-v3.5
• POST /v1/classify, POST /v1/segment, POST /v1/search with provider: "jina-search"
• Reader card still only serves /v1/web/fetch

[diegosouzapw]

Thanks for the thorough Jina + Gemini Embedding 2 multimodal completion -- the design (native pass-through to Jina, dashboard-key-first with env fallback, Gemini N-item batching) is solid and well tested. Two things need attention before merge: (1) the branch has drifted against the current release tip in open-sse/handlers/embeddings.ts, open-sse/handlers/search.ts and src/lib/embeddings/service.ts (two other PRs landed there since you branched) -- please rebase; (2) npm run lint is failing because of a few any types this PR introduces in open-sse/handlers/search.ts and src/lib/providers/validation*.ts, which is an error-level rule in open-sse/. Separately, this PR and #10567 both independently create src/shared/validation/jinaNativeEmbeddingInput.ts to solve the same Jina-native-image-input problem with different approaches -- since this PR's version is broader (also covers Gemini native parts and content-group fusion) we'd like to land this one as the canonical implementation and close #10567 as superseded, unless you'd prefer the other order.