feat: add avatar chat and knowledge workflow

2026-07-23 17:21:49 +08:00
parent 501f548bcc
commit 2d9f26a6f0
15 changed files with 3635 additions and 0 deletions
--- a/digital-avatar-app/backend/routers/chat.py
+++ b/digital-avatar-app/backend/routers/chat.py
@@ -0,0 +1,205 @@
+import difflib
+import os
+import re
+import string
+from typing import Any, Callable
+
+import httpx
+from fastapi import APIRouter, Body, Depends, Header, HTTPException
+from pydantic import BaseModel, Field
+from sqlalchemy.orm import Session
+
+import embeddings
+from database import get_db
+from models import Avatar, KnowledgeChunk, KnowledgeDoc, QAPair, User
+from responses import ok, fail
+
+router = APIRouter(tags=["数字分身聊天"])
+
+CHAT_API_URL = os.getenv("CHAT_API_URL", "https://dashscope.aliyuncs.com/compatible-mode/v1")
+CHAT_API_KEY = os.getenv("CHAT_API_KEY", "")
+CHAT_MODEL = os.getenv("CHAT_MODEL", "qwen-plus")
+MAX_MESSAGE_LENGTH = 4000
+MAX_HISTORY_MESSAGES = 10
+QA_SIMILARITY_THRESHOLD = 0.86
+
+
+class ChatMessage(BaseModel):
+    role: str = Field(pattern="^(user|assistant)$")
+    content: str = Field(min_length=1, max_length=MAX_MESSAGE_LENGTH)
+
+
+class ChatIn(BaseModel):
+    message: str = Field(min_length=1, max_length=MAX_MESSAGE_LENGTH)
+    history: list[ChatMessage] = Field(default_factory=list, max_length=MAX_HISTORY_MESSAGES)
+
+
+def _resolve_user(authorization: str | None, db: Session):
+    if not authorization:
+        return None
+    token = authorization.replace("Bearer ", "", 1).replace("bearer ", "", 1).strip()
+    return db.query(User).filter(User.app_token == token).first()
+
+
+def _require_owned_avatar(db: Session, avatar_id: str, authorization: str | None):
+    avatar = db.query(Avatar).filter(Avatar.id == avatar_id).first()
+    if not avatar:
+        raise HTTPException(status_code=404, detail="分身不存在")
+    user = _resolve_user(authorization, db)
+    if not user:
+        raise HTTPException(status_code=401, detail="未登录")
+    if avatar.owner_id and avatar.owner_id != user.huihui_user_id:
+        raise HTTPException(status_code=403, detail="无权访问该分身")
+    return avatar
+
+
+def _normalize_question(value: str) -> str:
+    value = (value or "").strip().lower()
+    value = re.sub(r"\s+", "", value)
+    return value.translate(str.maketrans("", "", string.punctuation + "，。！？；：、（）【】「」‘’“”《》"))
+
+
+def _match_standard_qa(question: str, qa_pairs: list[Any]):
+    normalized = _normalize_question(question)
+    if not normalized:
+        return None
+    enabled = [qa for qa in qa_pairs if getattr(qa, "enabled", True)]
+    for qa in enabled:
+        if _normalize_question(getattr(qa, "question", "")) == normalized:
+            return qa
+    best = None
+    best_score = 0.0
+    for qa in enabled:
+        candidate = _normalize_question(getattr(qa, "question", ""))
+        if not candidate:
+            continue
+        score = difflib.SequenceMatcher(None, normalized, candidate).ratio()
+        if score > best_score:
+            best, best_score = qa, score
+    return best if best_score >= QA_SIMILARITY_THRESHOLD else None
+
+
+def _config(avatar: Avatar) -> dict:
+    config = getattr(avatar, "config", None) or {}
+    return {
+        "replyStyle": config.get("replyStyle", "professional"),
+        "creativity": max(0, min(100, int(config.get("creativity", 50)))),
+        "rigor": max(0, min(100, int(config.get("rigor", 50)))),
+        "humor": max(0, min(100, int(config.get("humor", 30)))),
+        "responseLength": config.get("responseLength", "medium"),
+        "systemPrompt": (config.get("systemPrompt", "") or "").strip(),
+    }
+
+
+def _build_prompt(avatar: Avatar, history: list[Any], question: str, knowledge_hits: list[dict]) -> list[dict]:
+    config = _config(avatar)
+    knowledge = "\n".join(
+        f"[{hit.get('filename', '知识库')}] {hit.get('snippet', '')}"
+        for hit in knowledge_hits
+        if hit.get("snippet")
+    )
+    system = (
+        "你是用户的专属数字分身。请基于已提供的知识库回答，不要编造事实；"
+        f"回复风格：{config['replyStyle']}；严谨度：{config['rigor']}/100；"
+        f"幽默感：{config['humor']}/100；回复长度：{config['responseLength']}。"
+    )
+    if config["systemPrompt"]:
+        system += f"\n额外系统提示词：{config['systemPrompt']}"
+    if knowledge:
+        system += f"\n以下是可参考的知识库内容：\n{knowledge}"
+    messages = [{"role": "system", "content": system}]
+    for item in history[-MAX_HISTORY_MESSAGES:]:
+        messages.append({"role": item.role, "content": item.content} if hasattr(item, "role") else item)
+    messages.append({"role": "user", "content": question.strip()})
+    return messages
+
+
+def _search_knowledge(db: Session, avatar_id: str, question: str, top_k: int = 5) -> list[dict]:
+    chunks = db.query(KnowledgeChunk).filter(KnowledgeChunk.avatar_id == avatar_id).all()
+    if not chunks:
+        return []
+    qvec = embeddings.embed([question])[0]
+    scored = []
+    for chunk in chunks:
+        try:
+            vector = __import__("json").loads(chunk.vector)
+        except Exception:
+            continue
+        scored.append((embeddings.cosine(qvec, vector), chunk))
+    scored.sort(key=lambda item: item[0], reverse=True)
+    results = []
+    for score, chunk in scored[: max(1, top_k)]:
+        doc = db.query(KnowledgeDoc).filter(KnowledgeDoc.id == chunk.doc_id).first()
+        results.append({
+            "docId": chunk.doc_id,
+            "filename": doc.filename if doc else "",
+            "fileType": doc.file_type if doc else "",
+            "snippet": chunk.content[:120] + ("…" if len(chunk.content) > 120 else ""),
+            "score": round(score, 4),
+        })
+    return results
+
+
+def _call_qwen(messages: list[dict], temperature: float) -> str:
+    if not CHAT_API_KEY:
+        raise RuntimeError("Qwen 模型服务未配置 CHAT_API_KEY")
+    url = f"{CHAT_API_URL.rstrip('/')}/chat/completions"
+    payload = {
+        "model": CHAT_MODEL,
+        "messages": messages,
+        "temperature": temperature,
+    }
+    try:
+        response = httpx.post(
+            url,
+            headers={"Authorization": f"Bearer {CHAT_API_KEY}"},
+            json=payload,
+            timeout=30,
+        )
+        response.raise_for_status()
+        data = response.json()
+        answer = data.get("choices", [{}])[0].get("message", {}).get("content", "")
+    except (httpx.HTTPError, ValueError, KeyError, IndexError) as exc:
+        raise RuntimeError("Qwen 模型服务暂时不可用") from exc
+    if not isinstance(answer, str) or not answer.strip():
+        raise RuntimeError("Qwen 模型没有返回有效回答")
+    return answer.strip()
+
+
+def _resolve_reply(
+    db: Session,
+    avatar: Avatar,
+    question: str,
+    history: list[Any],
+    *,
+    qa_pairs: list[Any] | None = None,
+    search_fn: Callable[..., list[dict]] | None = None,
+    model_client: Callable[..., str] | None = None,
+) -> dict:
+    if qa_pairs is None:
+        qa_pairs = db.query(QAPair).filter(QAPair.avatar_id == avatar.id).all()
+    matched = _match_standard_qa(question, qa_pairs)
+    if matched:
+        return {"answer": matched.answer, "source": "qa", "references": []}
+
+    search_fn = search_fn or (lambda query, avatar_id: _search_knowledge(db, avatar_id, query))
+    hits = search_fn(question, avatar.id)
+    messages = _build_prompt(avatar, history, question, hits)
+    config = _config(avatar)
+    temperature = 0.2 + config["creativity"] / 100 * 0.6
+    model_client = model_client or _call_qwen
+    answer = model_client(messages=messages, temperature=temperature)
+    return {
+        "answer": answer,
+        "source": "knowledge" if hits else "qwen",
+        "references": hits,
+    }
+
+
+@router.post("/avatar/{avatar_id}/chat")
+def chat(avatar_id: str, body: ChatIn = Body(...), authorization: str = Header(None), db: Session = Depends(get_db)):
+    avatar = _require_owned_avatar(db, avatar_id, authorization)
+    try:
+        return ok(_resolve_reply(db, avatar, body.message, body.history))
+    except RuntimeError as exc:
+        return fail(str(exc), code=502)