Fix Replicate API for Gemini models and improve embed()

[Fuzzwah]

Summary

Fixes Replicate API to properly support Gemini models and improves the embed() function to be more robust.

Problem

When using Gemini models (e.g., google/gemini-2.5-flash) on Replicate:

1. The system_prompt field is ignored - Gemini needs the system message in the main prompt
2. The stream() method returns empty results - Gemini needs run() instead
3. The embed() function would incorrectly use the chat model for embeddings

Changes

Gemini Model Support

• Detect Gemini models by checking if model name contains 'gemini'
• Combine system message into the main prompt (since system_prompt is ignored)
• Use run() instead of stream() for Gemini models
• Handle various output formats (string, array, object)

Improved embed() Function

• Always use a dedicated embedding model, not the chat model
• Detect if configured model is an embedding model (contains 'embed', 'gte', 'e5-')
• Fall back to mark3labs/embeddings-gte-base for chat models
• Add input validation for text parameter
• Handle multiple output formats (vectors, embedding, embeddings, array)
• Better error messages

Usage

To use Gemini on Replicate, configure your profile:

{
  "model": "replicate/google/gemini-2.5-flash",
  "embedding": "replicate/mark3labs/embeddings-gte-base"
}

Testing

Tested with:

• google/gemini-2.5-flash for chat
• mark3labs/embeddings-gte-base for embeddings

Both work correctly with this fix.

[Sweaterdog]

Code looks good, I haven't tested