Model specs¶

Wherever a model is configured (server settings, model= with an explicit provider), ReadAnyDoc uses one string format:

<adapter>:<model>[#<effort>][?<json options>]

Part	Example	Meaning
adapter	`anthropic_api`	which provider integration to use
model	`claude-haiku-5-5`	the provider's model name
`#effort`	`#high`	effort level, for adapters that support it
`?json`	`?{"max_tokens":32000}`	extra request fields, sent as given

Adapters¶

Adapter	Provider	Effort	Key
`anthropic_api`	Anthropic Messages API	yes	`ANTHROPIC_API_KEY`
`anthropic_batch`	Anthropic Message Batches (half price, slow; for jobs that can wait)	yes	`ANTHROPIC_API_KEY`
`openai`	OpenAI API (or `OPENAI_BASE_URL`)	yes (`reasoning_effort`)	`OPENAI_API_KEY`
`openai_compatible`	any OpenAI-compatible server: `openai_compatible:<model>@<base_url>`	no	none

In the Python library you can write bare names: claude-* (and haiku, sonnet, opus) map to anthropic_api; gpt-*, o1, o3, o4, chatgpt-* map to openai. Anything else needs <adapter>:.

Examples¶

anthropic_api:claude-haiku-5-5#high                     the default page reader
anthropic_api:claude-sonnet-5-5                         the default handwriting reader
anthropic_api:claude-haiku-5-5#high?{"max_tokens":32000}
openai:gpt-5#medium
openai_compatible:Qwen/Qwen2.5-VL-7B-Instruct@http://localhost:8001/v1

Reserved options¶

model, messages, system, stream, tools, tool_choice, prompt and input are set by ReadAnyDoc and refused in options with a ModelSpecError.