Pipeline - AiQL Handbook

Pipeline language models run after convert. They do not read page pixels. They work on the cleaned text that layout, OCR, speech-to-text, and format converters already produced. They power shape and extract on the Middleware API. Split stays structure-aware (sections and headings) and is not this model family.

Parameters

Pipeline models are a sparse mixture-of-experts language model with a vision encoder. Shape and extract run on converted text. The same family can describe figure crops during convert.

Parameter	Value
Kind	Sparse mixture-of-experts, multimodal
Parameters	35 billion total, 3 billion active per token
Context	262k tokens native
Output budget	Up to 32k tokens per call

Configuration

Setting	How it is configured
Shape	Infer classes, properties, and relationships from the converted text.
Customer schema	A caller can supply a schema and skip discovery. Extract then follows that schema.
Extract	Pull entities and relationships as structured records guided by the schema.
Scope	One managed ingestion runs convert, split, shape, and extract. Each step is also a standalone tool.
Output length	Generation is sized for long extraction, not a short chat reply.
Serving	AiQL selects the checkpoint. Customers configure the job (source, optional schema), not the model name.