vllm_omni.protocol.realtime.items ¶

Realtime conversation items: shape, truncation and the transcript inside one.

A Realtime conversation is a list of items --- a user message, an assistant message, a function call --- each with a content list of typed parts. The client creates them (conversation.item.create), truncates an assistant item's audio after a barge-in (conversation.item.truncate), and the server echoes them back inside conversation.item.* and response.done events.

The rules in this module are the ones both sides have to agree on: which defaults a bare item gets, how far a transcript is cut when the client says it only heard N milliseconds of audio, and what makes a truncate request invalid. Item storage --- which items a session currently holds, in what order --- is the consumer's state, not this module's.

validate_realtime_video_frames sits here too: camera frames ride along an input_audio_buffer.append in the vLLM-Omni extension of the protocol, and validating them is the same kind of wire-shape check.

input_transcript_from_item(
    item: Mapping[str, object],
) -> str

normalize_conversation_item(
    item: Mapping[str, object],
) -> dict[str, object]

text_chars_for_audio_ms_from_marks(
    audio_end_ms: int,
    text_len: int,
    marks: list[object],
    *,
    final_ms: object | None = None,
) -> int

truncate_realtime_item_content(
    item: dict[str, object],
    *,
    content_index: int,
    audio_end_ms: int,
) -> None

validate_realtime_item_truncate(
    item: Mapping[str, object],
    *,
    content_index: int,
    audio_end_ms: int,
) -> str | None

validate_realtime_video_frames(
    video_frames: object, max_slice_nums: object
) -> str | None