[BUG] OpenAI Embeddings Custom silently falls back to 192-dim embeddings when using LiteLLM + Infinity backend

[periti1]

When using the "OpenAI Embeddings Custom" node connected to a LiteLLM proxy with Infinity as the embedding backend, Flowise silently falls back to an internal 192-dimension embedder instead of using the
configured API.

This causes a dimension mismatch error when querying vector stores that were populated with the correct 768-dimension embeddings:
Error searching for documents: 22000 different vector dimensions 768 and 192 null

To Reproduce

1. Set up LiteLLM proxy with Infinity backend serving nomic-embed-text (768 dims)
2. Create a flow with "OpenAI Embeddings Custom" node:
   • BasePath: http://<litellm-host>:4000/v1
   • Model: nomic-embed-text
   • Credential: Valid API key
3. Connect to a Supabase/pgvector store with 768-dim vectors
4. Try to query the vector store

Expected behavior

Flowise should either:

1. Successfully call the LiteLLM/Infinity API and return 768-dim embeddings, OR
2. Throw a clear error if the API call fails

Current behavior

• Flowise silently falls back to an internal 192-dimension embedder
• No error is shown in the UI
• The vector search fails due to dimension mismatch

Evidence that the API works correctly

Calling the same endpoint from CLI returns correct 768-dim embeddings:

curl -s http://<litellm-host>:4000/v1/embeddings \
  -H "Authorization: Bearer <api-key>" \
  -H "Content-Type: application/json" \
  -d '{"model": "nomic-embed-text", "input": "test"}'
# Returns embedding with 768 dimensions ✓

Workaround

Using the native "Ollama Embeddings" node instead of "OpenAI Embeddings Custom" works correctly with the same model.

Environment

- Flowise version: latest (Docker image flowiseai/flowise:latest)
- LiteLLM version: latest
- Infinity version: latest
- Embedding model: nomic-embed-text (768 dimensions)
- Vector store: Supabase with pgvector

Additional context

The issue appears to be in how Flowise handles the response from LiteLLM/Infinity. The fallback to 192 dimensions happens silently without any error logging, making it very difficult to debug.