Function answer_question_stream

agents/wiki-agent/retrieve/answer.py:256–318

Yield answer events with local citations and provider-native chunks.

(
    user_id: str,
    question: str,
    *,
    limit: int = 8,
    pinned_nodes: Optional[List[Dict[str, Any]]] = None,
    allowed_node_ids: Optional[Iterable[str]] = None,
    response_language: str = "zh",
)

def answer_question_stream(
    user_id: str,
    question: str,
    *,
    limit: int = 8,
    pinned_nodes: Optional[List[Dict[str, Any]]] = None,
    allowed_node_ids: Optional[Iterable[str]] = None,
    response_language: str = "zh",
) -> Iterator[Dict[str, Any]]:
    """Yield answer events with local citations and provider-native chunks."""
    language = _normalize_response_language(response_language)
    started = time.time()
    hits, embedding_error = _prepare_hits(user_id, question, limit, pinned_nodes, allowed_node_ids=allowed_node_ids)
    if not hits:
        result = _empty_answer(started, response_language=language)
        yield {"event": "meta", "data": {key: value for key, value in result.items() if key != "text"}}
        yield {"event": "chunk", "data": {"text": result["text"]}}
        yield {"event": "done", "data": result}
        return

    citations = _build_citations(user_id, hits)
    llm = build_llm_provider()
    prompt = _build_prompt(question, hits)
    llm_error = None
    text_parts: List[str] = []
    meta = {
        "citations": citations,
        "elapsed_ms": 0,
        "response_language": language,
        "token_usage": _token_usage(llm, None, embedding_error, None),
        "streaming": {"provider": True, "transport": "sse"},
    }
    yield {"event": "meta", "data": meta}
    try:
        for chunk in llm.stream_generate(prompt, system=_system_prompt(language), temperature=0.0, max_tokens=_answer_max_tokens()):
            if not chunk:
                continue
            text_parts.append(chunk)
            yield {"event": "chunk", "data": {"text": chunk}}
    except Exception as exc:
        llm_error = str(exc)
        fallback = _extractive_fallback(hits, llm_error, response_language=language)
        text_parts = [fallback]
        yield {"event": "chunk", "data": {"text": fallback}}

    answer_text = "".join(text_parts)
    response = None
    if llm_error is None and _looks_incomplete_answer(answer_text):
        try:
            response = llm.generate(prompt, system=_system_prompt(language), temperature=0.0, max_tokens=_answer_max_tokens())
            if response.text and len(response.text.strip()) > len(answer_text.strip()):
                answer_text = response.text
        except Exception as exc:
            llm_error = f"incomplete stream fallback failed: {exc}"

    result = {
        "text": answer_text,
        "citations": citations,