feat: add Gemini embeddings support + fix missing OpenAI response fields (Letta compat)

[xuandung38]

Summary

Two improvements to 9router:

1. Gemini embeddings support — extend the existing /v1/embeddings endpoint with Google AI (Gemini) provider
2. Fix: inject missing object/created fields — makes responses compatible with strict OpenAI clients

---

1. Gemini Embeddings Support

The existing /v1/embeddings endpoint (added in #146) only supports OpenAI passthrough. This PR adds Google AI (Gemini) as an additional embeddings provider.

Why

Many use cases require embeddings that are either free-tier or do not depend on a paid OpenAI subscription. Gemini offers high-quality embeddings (up to 3072 dimensions) with a generous free quota.

⚠️ Provider support at this time:

• ✅ Google AI (Gemini) — fully supported in this PR
• ✅ OpenAI — supported via passthrough (existing)
• ❌ GitHub Copilot, Anthropic/Claude, and other providers do not have an embeddings API and are not supported

New models

Model	Dimensions	Notes
gemini/gemini-embedding-001	3072	Latest, highest quality
gemini/text-embedding-005	768	Efficient
gemini/text-embedding-004	768	Stable

Setup

Add a Google AI provider in your 9router provider config with a valid Gemini API key.
Get a free key at: https://aistudio.google.com/app/apikey

Usage

# Gemini embeddings
curl http://localhost:20128/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{"model": "gemini/gemini-embedding-001", "input": "Hello world"}'

# OpenAI embeddings (existing, unchanged)
curl http://localhost:20128/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{"model": "openai/text-embedding-3-small", "input": "Hello world"}'

Response follows standard OpenAI format:

{
  "object": "list",
  "data": [{"object": "embedding", "index": 0, "embedding": [0.123, ...]}],
  "model": "gemini-embedding-001",
  "usage": {"prompt_tokens": 2, "total_tokens": 2}
}

---

2. Fix: Inject Missing object / created Fields

Why

Some upstream providers (notably GitHub Copilot) return chat completion responses without the object and created fields. This causes strict OpenAI-compatible clients to reject the response with a validation error.

One concrete example: Letta (open-source memory layer) validates every LLM response against the OpenAI ChatCompletionResponse schema using Pydantic. A missing created or object field results in a 422 error, making it impossible to use 9router as Letta's LLM backend.

Fix

After receiving a response from the upstream provider, inject the missing fields if absent:

• Non-streaming: "object": "chat.completion" and "created": <unix_timestamp>
• Streaming chunks: "object": "chat.completion.chunk" and "created": <unix_timestamp>

This is a safe, additive change — providers that already include these fields are unaffected.

---

Files Changed

File	Change
open-sse/services/embeddingsCore.js	Gemini provider: different URL, query-param auth, normalized response
open-sse/config/providerModels.js	Register 3 Gemini embedding models
open-sse/handlers/chatCore.js	Inject object/created for non-streaming responses
open-sse/utils/stream.js	Inject object/created for each SSE chunk (streaming passthrough)

Breaking Changes

None.

Testing

• Gemini gemini-embedding-001 → 3072-dim vectors ✅
• Letta integration: confirmed working end-to-end after the object/created fix ✅

[xuandung38]

WIP: please wait 🗡️

[xuandung38]

Ready for review and merge

[decolua]

Cherry-picked commits 1, 2, 5, 6 into our fork. Thanks @xuandung38! 🙏