nianjie/dialog/backend/routes/chat.py

import json
import re
import time
import uuid
from datetime import datetime

from flask import Blueprint, request, jsonify, Response, stream_with_context
from openai import OpenAI

from ..config import TOKEN_INTERVAL, MODEL_NAME, OPENAI_BASE_URL, API_KEY
from ..prompt import SYSTEM_PROMPT
from ..rag import search_rag
from ..conversation_store import make_conversation, load as load_conversation, save as save_conversation

bp = Blueprint("chat", __name__, url_prefix="/api")

client = OpenAI(api_key=API_KEY, base_url=OPENAI_BASE_URL)

TOOLS = [
    {
        "type": "function",
        "function": {
            "name": "search_rag",
            "description": "在本地 QA 数据集中按 question 字段检索，返回最多 5 条匹配。",
            "parameters": {
                "type": "object",
                "required": ["query"],
                "properties": {
                    "query": {
                        "type": "string",
                        "description": "用户问题或关键词"
                    }
                }
            }
        }
    }
]


def sanitize_legacy_markers(text: str) -> str:
    if not text:
        return text
    text = text.replace("[[SEARCH_START]]搜索中...", "")
    text = text.replace("[[SEARCH_DONE]]搜索完成", "")
    text = text.replace("[[SEARCH_START]]", "")
    text = text.replace("[[SEARCH_DONE]]", "")
    text = re.sub(r"(?m)^[ \t]*搜索中\.\.\.[ \t]*\n?", "", text)
    text = re.sub(r"(?m)^[ \t]*搜索完成[ \t]*\n?", "", text)
    return text


def messages_for_model(messages: list[dict]) -> list[dict]:
    sanitized = []
    for m in messages:
        mm = dict(m)
        if mm.get("role") == "assistant" and isinstance(mm.get("content"), str):
            mm["content"] = sanitize_legacy_markers(mm["content"])
        sanitized.append(mm)
    return sanitized


def emit_events_for_text(text: str):
    for ch in text:
        yield json.dumps({"type": "assistant_delta", "delta": ch}, ensure_ascii=False) + "\n"
        time.sleep(TOKEN_INTERVAL)


@bp.post("/chat")
def chat():
    data = request.get_json(force=True)
    user_text = (data.get('message') or '').strip()
    if not user_text:
        return jsonify({"error": "message is required"}), 400

    cid = data.get('conversation_id')
    convo = load_conversation(cid) if cid else None
    if not convo:
        convo = make_conversation(user_text[:20])
        cid = convo['id']

    # 仅记录 AI 对话的消息
    convo['messages'].append({"role": "user", "content": user_text})
    convo['updated_at'] = datetime.utcnow().isoformat()
    save_conversation(convo)

    all_messages = [SYSTEM_PROMPT] + messages_for_model(convo['messages'])

    def generate():
        new_messages = []
        try:
            model_messages = list(all_messages)
            max_tool_rounds = 2
            tool_round = 0

            while True:
                completion = client.chat.completions.create(
                    model=MODEL_NAME,
                    messages=model_messages,
                    temperature=0.6,
                    stream=True,
                    tools=TOOLS,
                    tool_choice="auto",
                )
                print("[chat] streaming start, cid=", cid, "tool_round=", tool_round, "len(messages)=", len(model_messages))

                tool_calls_acc = {}
                segment_text = ""
                tool_start_emitted = False

                for chunk in completion:
                    delta = chunk.choices[0].delta
                    text = delta.content or ""
                    if text:
                        segment_text += text
                        yield from emit_events_for_text(text)
                        print("[chat] delta text len", len(text))

                    if delta.tool_calls:
                        if not tool_start_emitted:
                            tool_start_emitted = True
                            yield json.dumps({"type": "tool_call_start"}, ensure_ascii=False) + "\n"
                            print("[chat] tool_call_start")
                        for tc in delta.tool_calls:
                            idx = tc.index or 0
                            entry = tool_calls_acc.setdefault(idx, {"id": tc.id, "name": None, "arguments": ""})
                            if tc.id:
                                entry["id"] = tc.id
                            if tc.function:
                                if tc.function.name:
                                    entry["name"] = tc.function.name
                                if tc.function.arguments:
                                    entry["arguments"] += tc.function.arguments
                                    print("[chat] tool_call accumulating args len", len(entry["arguments"]))

                if not tool_calls_acc:
                    if segment_text:
                        new_messages.append({"role": "assistant", "content": segment_text})
                    break

                if tool_round >= max_tool_rounds:
                    if segment_text:
                        new_messages.append({"role": "assistant", "content": segment_text})
                    break

                tool_round += 1

                call = list(tool_calls_acc.values())[0]
                tool_name = call.get("name")
                tool_args = call.get("arguments") or ""
                tool_call_id = call.get("id") or str(uuid.uuid4())

                yield json.dumps(
                    {
                        "type": "tool_call",
                        "tool_call_id": tool_call_id,
                        "name": tool_name,
                        "arguments": tool_args,
                    },
                    ensure_ascii=False,
                ) + "\n"
                print("[chat] tool_call emit", tool_name, tool_args)

                query = ""
                try:
                    parsed = json.loads(tool_args or "{}")
                    query = parsed.get("query", "")
                except Exception:
                    query = tool_args

                matches = search_rag(query, limit=5)
                tool_response_content = json.dumps({"query": query, "matches": matches}, ensure_ascii=False)

                time.sleep(0.5)  # 缩短等待，便于前端即时显示

                yield json.dumps(
                    {
                        "type": "tool_result",
                        "tool_call_id": tool_call_id,
                        "content": tool_response_content,
                    },
                    ensure_ascii=False,
                ) + "\n"
                print("[chat] tool_result emit, matches", len(matches))

                assistant_tool_msg = {
                    "role": "assistant",
                    "content": segment_text,
                    "tool_calls": [{
                        "id": tool_call_id,
                        "type": "function",
                        "function": {"name": tool_name, "arguments": tool_args}
                    }]
                }
                tool_result_msg = {
                    "role": "tool",
                    "tool_call_id": tool_call_id,
                    "content": tool_response_content,
                }

                model_messages = model_messages + [assistant_tool_msg, tool_result_msg]
                new_messages.append(assistant_tool_msg)
                new_messages.append(tool_result_msg)

        except Exception as e:
            err = f"[出错]{e}"
            yield from emit_events_for_text(err)
            new_messages.append({"role": "assistant", "content": err})
        finally:
            if new_messages:
                convo['messages'].extend(new_messages)
                convo['updated_at'] = datetime.utcnow().isoformat()
                save_conversation(convo)

    headers = {'X-Conversation-Id': cid}
    return Response(stream_with_context(generate()), mimetype='application/x-ndjson; charset=utf-8', headers=headers)