LLM Clients

AI-Parrot talks to every LLM provider through a single AbstractClient interface, so the same agent code can run on OpenAI, Anthropic, Google GenAI, Groq, VertexAI or HuggingFace without changes.

What lives here

• AbstractClient — async base class with completion(), stream() and embed(). Every provider wrapper inherits from it.
• Provider clients — concrete implementations under parrot.clients.* (one module per vendor).
• LLM_PRESETS — opinionated configurations for the most common models (temperature, max tokens, retry budget).
• StreamingRetryConfig — retry policy applied around streaming completions, including back-off and partial-response handling.
• Embeddings — parrot.embeddings covers dense and sparse embedding back-ends used by the RAG pipeline.

flowchart LR
    Agent --> AbstractClient
    AbstractClient --> OpenAI
    AbstractClient --> Anthropic
    AbstractClient --> Groq
    AbstractClient --> VertexAI
    AbstractClient --> HuggingFace

Golden rules

1. Never call a provider SDK directly — always go through AbstractClient. This is the seam that keeps the framework vendor agnostic.
2. await everything — clients are async only. Wrapping them with asyncio.run inside another coroutine will block the event loop.
3. Don't subclass AbstractClient lightly — discuss the change first; it's the foundation that all bots and crews assume.

Read next

• Per-loop Caching — how the client layer caches completions inside a single agent turn.
• Contextual Embedding
• Matryoshka Embeddings

API reference

API Reference → Clients