Jun 2025 (v0.3.2)

Mochi v0.3.2 introduces model aliases and embedding generation, expanding the language’s integration with LLM-backed workflows. These additions make it easier to manage AI providers and extract semantic representations of text—supporting more advanced applications in ranking, similarity, and hybrid reasoning.

Model Aliases

Mochi programs can now define reusable model configurations using model blocks. This allows prompts and generation logic to remain clean and declarative, while model settings are specified once and referenced by name.

model fast {
  provider: "openai"
  name: "gpt-3.5-turbo"
}

let response = generate text {
  model: "fast"
  prompt: "Summarize this week’s release notes"
}

Model aliases apply to all forms of generation (text, embedding, structured) and delegate to the shared runtime backend.

Embedding Generation

Mochi now supports generating vector embeddings from text using generate embedding blocks. The result is a list<float> suitable for downstream semantic tasks.

let vec = generate embedding {
  text: "hello world"
  normalize: true
}

print(len(vec))

Normalization ensures the resulting vector has unit length, which is useful for cosine-based comparisons.

Additional Updates

• All generate forms now route through a unified backend with simplified behavior
• Legacy interpolation logic inside generate blocks has been removed
• Internal test coverage has been expanded for embedding and alias support