Documentation

Complete guide to using UnisAI for switching between providers or models. Learn how to migrate prompts, embeddings, and training data formats.

🚀Quick Start

🚀

Get Started in 60 Seconds

Try our interactive demo with sample data, or start migrating your own prompts right away.

Quick Steps:
1. Select source & target providers
2. Choose model type (Chat / Embedding / Fine-tune)
3. Enter your prompt or data
4. Click "Run Migration"
5. Get your converted output instantly!
API quickstart (one prompt, two providers)

API base for this deployment: https://ai-migrator-backend-production.up.railway.app

Copy the curl below—it uses the live API above. Add your provider API keys in the request body if you want live validation.

curl -X POST https://ai-migrator-backend-production.up.railway.app/api/migrate/run \
  -H "Content-Type: application/json" \
  -d '{
    "source_provider": "openai",
    "target_provider": "anthropic",
    "model_type": "chat",
    "prompt": "Summarize in one sentence."
  }'

Response includes conversion_score, cost_comparison, summary, and when API keys are set, metadata with output parity validation.

✨Enhanced Services

All conversion services include advanced quality metrics and optimization features. Compare features across services:

Feature
💬Prompt
🔢Embedding
🎯Fine-Tuning
Semantic Similarity Scoring✓ Embedding-based✓ Preservation metrics—
Structure/Format Preservation✓ Metrics included—✓ Format conversion
Quality Levels✓ Excellent/Good/Moderate✓ Based on method✓ Quality scoring
Optimization✓ Optional prompt optimization✓ Auto method selection✓ Hyperparameter recommendations
Real API Validation✓ When API keys configured✓ OpenAI & Voyage AI— Format only
Advanced Metrics✓ Multi-metric assessment✓ Variance preservation✓ Data quality analysis
Cost/Time Estimation——✓ Included
Method Compatibility—✓ PCA/SVD/Truncate✓ SFT/DPO/RFT/Vision
Recommendations✓ Actionable improvements✓ Based on sample size✓ Training data quality

💡 Note: Enhanced services are automatically enabled for all conversions. Quality metrics are calculated using advanced algorithms and, when API keys are configured, real API validation is performed.

⚙️How It Works

1

Input Your Data

Provide your prompts, embeddings, or training data from OpenAI or Anthropic

2

Select Target Provider

Choose where you want to migrate to (OpenAI, Anthropic, Voyage AI, or more coming soon)

3

Automatic Conversion

UnisAI converts your data while preserving semantic meaning

4

Use Your Converted Data

Get your converted data ready to use with the new provider

Scope: Today we focus on static prompt/template conversion and validation. We do not yet capture per-request dynamic context (e.g. retrieval context, function outputs, or full "replay what the LLM saw"). Checkpoints and lineage for dynamic context may be added in a future release.

🔌Supported Providers

OpenAI

✅ Format Conversion Supported

16 chat models (GPT-5.2, GPT-5.1, GPT-4.1, GPT-4o, o3, o4-mini) + 2 embedding + 7 fine-tune

16 chat, 2 embed, 7 fine-tune

Real API integration requires API keys

ChatEmbeddingsFine-tuning

Anthropic

✅ Format Conversion Supported

11 models (Claude Opus 4.1, Claude Sonnet 4, Claude 4.5, Claude 3.5, Claude 3)

11 chat, 1 fine-tune*

*Via AWS Bedrock (SFT only). Real API integration requires API keys

ChatFine-tuning*

Voyage AI

✅ Format Conversion Supported

7 embedding models (Anthropic recommended)

7 embedding

FREE: 200M tokens/month! Real API integration requires API key

Embeddings

Google

🚧 Coming Soon

Gemini Pro, Ultra

ChatEmbeddings

Meta LLaMA

🚧 Coming Soon

LLaMA 2, 3

ChatFine-tuning

We're currently focused on deep integration with OpenAI, Anthropic, and Voyage AI. Support for Google Gemini, Meta LLaMA, and other providers is coming soon.

📊 View Complete Model Support Table →

✨Features

❓Frequently Asked Questions

💡Best Practices

🛠️Additional Tools

🚀

Beta Access Available

Advanced tools including Benchmarking, Analytics Dashboard, and Embedding Aligner are available for beta testers. Join our waitlist to get early access to these powerful features.

Join Beta Waitlist→

🔧Troubleshooting

Migration fails with "Provider not supported"
Ensure you've selected a supported provider. Check the Supported Providers section above for the full list.
Embedding alignment quality is low
Try training with more data pairs. Use the Embedding Aligner Panel to test different alignment methods. Ensure your source and target embeddings are from the same domain.
Conversion takes too long
Large prompts or datasets may take longer. For embeddings, alignment training can take 10-30 seconds. Check your network connection and try again.
Not seeing real API results
Format conversion works without API keys, but real API validation requires API keys to be configured on the backend. Check the backend configuration - if DEMO_MODE is false and API keys are set, you'll get real API results. Otherwise, you'll see format conversion with mock data for validation.
Quality score is lower than expected
Quality scores vary based on conversion complexity. Format conversion (structure) maintains 95%+ accuracy. Semantic similarity for prompts typically ranges 70-85% (varies by use case). Embedding conversions achieve 85-95% with PCA/SVD and sufficient samples (100+). Fine-tune format conversions show 80-90% depending on model compatibility. Check the quality metrics breakdown for detailed analysis.

💬Need More Help?

Still have questions? Check the FAQ above or reach out to our team.