leaudit-platform-backend/fastapi_modules/fastapi_leaudit/rag_engine/generator.py

from __future__ import annotations

import json
import time
import uuid
from typing import AsyncGenerator

import httpx

from fastapi_modules.fastapi_leaudit.rag_engine.config import RAG_CONFIG, build_openai_chat_completions_url

DEFAULT_SYSTEM_PROMPT = """你是烟草行业智慧法务小助手，专注于烟草专卖法规、合同管理、行政处罚等相关法律法规。\n\n回答要求：\n- 先用一句话直接回答，再展开详细说明\n- 多个要点用编号列表\n- 关键法条和数字用 **加粗**\n- 分类信息用表格\n- 层级结构用缩进子列表\n- 不要加标题，直接输出正文"""


async def generate_stream(
    query: str,
    context_chunks: list[dict],
    conversation_id: str,
    message_id: str,
    task_id: str | None = None,
    system_prompt: str = "",
    model: str = "",
    temperature: float | None = None,
    max_tokens: int | None = None,
    dataset_name: str = "",
) -> AsyncGenerator[str, None]:
    task_id = task_id or str(uuid.uuid4())
    created_at = int(time.time())
    _model = model or RAG_CONFIG["LLM_MODEL"]
    _temp = temperature if temperature is not None else RAG_CONFIG["LLM_TEMPERATURE"]
    _max_tok = max_tokens or RAG_CONFIG["LLM_MAX_TOKENS"]
    _prompt = system_prompt or DEFAULT_SYSTEM_PROMPT

    max_context_chars = 8000
    if context_chunks:
        attachment_parts: list[str] = []
        formal_parts: list[str] = []
        total_len = 0
        for chunk in context_chunks:
            scope = str(chunk.get("source_scope") or chunk.get("data_source_type") or "formal_kb")
            if scope == "chat_attachment":
                label = "上传文件事实"
            else:
                label = "正式知识库依据"
            part = f"[{label} | 来源: {chunk.get('source', '未知')}]\\n{chunk.get('text', '')}"
            if total_len + len(part) > max_context_chars:
                break
            if scope == "chat_attachment":
                attachment_parts.append(part)
            else:
                formal_parts.append(part)
            total_len += len(part)
        context_sections: list[str] = []
        if attachment_parts:
            context_sections.append("【上传文件事实】\n" + "\\n\\n---\\n\\n".join(attachment_parts))
        if formal_parts:
            context_sections.append("【正式知识库依据】\n" + "\\n\\n---\\n\\n".join(formal_parts))
        context_text = "\\n\\n======\\n\\n".join(context_sections)
        user_content = (
            "请严格区分上下文来源：上传文件事实只能用于判断用户材料中的事实；正式知识库依据只能用于法律、处罚、裁量或案例依据。"
            "如果依据不足，请明确说明，不要编造。\n\n"
            f"{context_text}\\n\\n用户问题: {query}"
        )
    else:
        user_content = query

    messages = [
        {"role": "system", "content": _prompt},
        {"role": "user", "content": user_content},
    ]

    total_tokens = 0
    try:
        async with httpx.AsyncClient(timeout=RAG_CONFIG["LLM_TIMEOUT"]) as client:
            async with client.stream(
                "POST",
                build_openai_chat_completions_url(RAG_CONFIG["LLM_BASE_URL"]),
                json={
                    "model": _model,
                    "messages": messages,
                    "temperature": _temp,
                    "max_tokens": _max_tok,
                    "stream": True,
                    "chat_template_kwargs": {"enable_thinking": False},
                },
                headers={
                    "Content-Type": "application/json",
                    "Authorization": f"Bearer {RAG_CONFIG['LLM_API_KEY']}",
                },
            ) as resp:
                resp.raise_for_status()
                async for line in resp.aiter_lines():
                    if not line.startswith("data: "):
                        continue
                    payload = line[6:].strip()
                    if payload == "[DONE]":
                        break
                    chunk = json.loads(payload)
                    choices = chunk.get("choices") or []
                    first_choice = choices[0] if choices and isinstance(choices[0], dict) else {}
                    delta = first_choice.get("delta", {})
                    text = delta.get("content", "")
                    if text:
                        yield _sse_line(
                            {
                                "event": "message",
                                "task_id": task_id,
                                "message_id": message_id,
                                "conversation_id": conversation_id,
                                "answer": text,
                                "created_at": created_at,
                            }
                        )
                    usage = chunk.get("usage")
                    if usage:
                        total_tokens = usage.get("total_tokens", total_tokens)
    except Exception as exc:
        yield _sse_line(
            {
                "event": "error",
                "task_id": task_id,
                "message_id": message_id,
                "code": "llm_error",
                "message": str(exc),
            }
        )
        return

    retriever_resources = [
        {
            "position": i + 1,
            "dataset_id": "",
            "dataset_name": dataset_name,
            "document_id": "",
            "document_name": chunk.get("source", ""),
            "data_source_type": chunk.get("data_source_type") or chunk.get("source_scope") or "upload_file",
            "segment_id": chunk.get("id", ""),
            "retriever_from": "rag",
            "score": round(chunk.get("score", 0.0), 4),
            "hit_count": 0,
            "word_count": len(chunk.get("text", "")),
            "segment_position": i + 1,
            "index_node_hash": "",
            "content": chunk.get("text", "")[:500],
            "page": None,
            "source_scope": chunk.get("source_scope") or "",
            "attachment_id": chunk.get("attachment_id") or "",
        }
        for i, chunk in enumerate(context_chunks)
    ]

    yield _sse_line(
        {
            "event": "message_end",
            "task_id": task_id,
            "message_id": message_id,
            "conversation_id": conversation_id,
            "metadata": {
                "usage": {"total_tokens": total_tokens},
                "retriever_resources": retriever_resources,
            },
        }
    )


def _sse_line(data: dict) -> str:
    return f"data: {json.dumps(data, ensure_ascii=False)}\n\n"