GenAI Processors is a Python library for building generative AI pipelines and agents out of small, composable, asynchronous units. It rests on three ideas: a unified content model that represents inputs and outputs the same way across models, agents, and tools, processors that transform streams of that content using native asyncio, and streaming as the default rather than an add-on.
A Processor encapsulates one unit of work. Authors implement it as an async generator that consumes a stream of content parts and yields new parts, and callers can gather the whole result, read it as text, or iterate over parts as they arrive. Processors chain sequentially with + and run in parallel with //, so complex flows and agent behaviors are assembled from reusable pieces.
The same abstraction covers single Gemini model calls, turn-based conversations, and real-time streaming through the Gemini Live API. Examples include an audio-in, audio-out live agent with Google Search as a tool, a research agent built from three sub-processors, and a live commentary agent made of an event-detection agent and a conversation agent.
Features
- Unified content model: ProcessorPart wraps a Gemini content part with MIME type, role, and custom metadata, covering text, images, audio, and JSON
- Composable processors: chain units with + and parallelise them with // to build data flows and agent behaviors
- Streaming by default: results can be gathered whole, read as text, or consumed part by part as they stream
- Gemini integration: ready-made processors for turn-based model calls and real-time Live API streaming
- Custom processors: create new ones by subclassing the base classes or decorating plain functions
- Stream utilities: split, concatenate, and merge asynchronous streams of parts
- Concurrency: asyncio-based orchestration of network I/O and compute-heavy subthreads
- Built-in and contributed processors: a core set of building blocks for real-time applications plus community contributions
- Code generation instructions: an llms.txt file to steer coding models toward current library APIs
