You're iterating a streaming LLM response and want to know what chunk types you can get and how to extract literal text. I’m responding about the async loop and extracting text from

llm.ChatChunk

. Quick answer: • The most common yielded type is

llm.ChatChunk

. It can include: ◦ text deltas (partial/streamed text) — available at

chunk.delta.content

◦ tool calls (instructions to call an external tool) ◦ metadata (usage, reasoning, events) • You may also see raw strings or

None

in some implementations. To safely get the generated text, check for

ChatChunk

and

delta.content

, and fall back to stringifying other chunk types. Example pattern:

async for chunk in self._llm.chat(chat_ctx=chat_ctx):
    if isinstance(chunk, llm.ChatChunk) and hasattr(chunk.delta, "content"):
        text = chunk.delta.content
    else:
        text = str(chunk) if chunk is not None else None

    if text:
        # handle the text (print, yield, append, etc.)
        print(text, end="")

If you want, tell me: • which llm client/library/version you’re using, and • whether you need full responses assembled or incremental streaming output — I’ll give a tailored snippet. Sources: Modifying LLM output before TTS | LiveKit docs | Pipeline nodes and hooks | LiveKit docs