urai

Urai means “dialog” in Tamil. It defines the chat behaviour shared by every backend. aidialog defines message parts. A backend package such as rishi talks to an inference engine. Urai manages the conversation between those boundaries.

A backend implements one model step. Urai supplies model resolution, portable options, history, tools, streaming, callbacks, and usage accounting.

Install

pip install urai

Urai does not include an inference backend. Install one separately, then import its backend module to register a runtime.

Start a chat

The examples below assume a backend has registered a remote runtime. Pass its model name to mk_chat, then call the returned chat with a message:

from urai import mk_chat, resp_text

chat = mk_chat('gpt-5.1', runtime='remote', sp='Answer briefly.', temp=0.2)
reply = chat('What is 6 times 7?')
print(resp_text(reply))

chat.hist contains the normalized conversation. The same history shape works with every backend.

[(msg['role'], resp_text(msg)) for msg in chat.hist]

Resolve a model before loading it

resolve identifies a registered runtime and context window without building a backend:

from urai import resolve

spec = resolve('claude-sonnet-4.6', runtime='remote')
print(spec.runtime, spec.ctx, spec.local)

chat = mk_chat(spec)

A resolved ModelSpec can be cached, compared, and passed between layers.

Configure one turn

Constructor options set defaults for the conversation. Generation options passed to a call apply only to that turn:

reply = chat('Explain the hard part.', effort='high', temp=0.9)

Urai translates portable names such as ctx and temp to backend-specific names. Unsupported options produce a warning. Backend-only options belong in extra.

Add a tool

A Python callable becomes a tool from its signature and docstring:

def add(a: int, b: int) -> int:
    "Add two integers."
    return a + b

chat = mk_chat('gpt-5.1', runtime='remote', tools=[add], approve=lambda call: True)
reply = chat('Use the tool to add 20 and 22.')
print(resp_text(reply))

Approval runs before each local tool call. Provider-run tools are recorded but never executed locally.

Modules

Write a backend

Subclass Chat, declare option translations, and register the runtime. Add ToolLoopMixin when the backend returns tool calls as data.

from urai import Chat, ToolLoopMixin, Runtime, register_runtime

class MyChat(ToolLoopMixin, Chat):
    _runtime = 'mine'
    _opt_map = {'ctx': 'n_ctx'}
    _opt_skip = ('effort',)

    def _model_step(self, **kw): ...
    def _stream_step(self, **kw): ...

register_runtime(Runtime('mine', 'mypkg.chat.MyChat', pats=('.mine',)))

Chat('model.mine') now dispatches to MyChat.

Develop

The notebooks in nbs/ are the source. Files in urai/ are generated.

uv sync --all-extras --group dev
uv run nbdev-export
uv run nbdev-test
uv run nbdev-clean

The test suite uses scripted backends. It does not load a model or require network access.