feat(agent): 智能体接入工具调用 + 园区知识库（RAG）
后端： - tools.py：工具注册表（get_park_overview 园区实时数据 / query_companies 企业名录 / control_display 大屏控制 / get_time），POST /api/tools/exec 执行 - rag.py：园区知识库（park.md 分块 → dashscope text-embedding-v3 → numpy 余弦检索），/api/kb/brief 摘要注入 + /api/kb/retrieve 动态检索 - 知识索引 kb_index.json 入库，部署免首次构建 前端（VoiceAssistant）： - TOOL_DEFS 经 session.update 传给 LLM；toolcall 事件 → 后端执行 → sendToolOutput + requestResponse - instructions 拼接园区知识摘要；工具调用气泡提示 - 测试通过：4 个工具，检索「入驻政策」正确命中政策段落
This commit is contained in:
Pine
@@ -0,0 +1,102 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""园区知识库 —— 轻量本地 RAG（embedding + numpy 余弦检索）
|
||||
|
||||
- 知识源：backend/knowledge/park.md（按标题/段落分块）
|
||||
- 向量化：dashscope text-embedding-v3（云端 API，索引构建一次后缓存）
|
||||
- 检索：numpy 余弦 top-k（资料量小，暴力检索足够）
|
||||
- 提供：
|
||||
- retrieve(query, top_k) → 相关段落（供对话动态注入）
|
||||
- brief() → 固定知识摘要（供前端拼进 instructions，静态知识一次注入）
|
||||
"""
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
import numpy as np
|
||||
|
||||
log = logging.getLogger("dpm.rag")
|
||||
|
||||
KB_FILE = Path(__file__).resolve().parent.parent / "knowledge" / "park.md"
|
||||
INDEX_FILE = Path(__file__).resolve().parent.parent / "knowledge" / "kb_index.json"
|
||||
EMBED_MODEL = "text-embedding-v3"
|
||||
|
||||
_index = None # 缓存 {chunks, embeddings}
|
||||
|
||||
|
||||
def _chunk_md(text: str) -> list[str]:
|
||||
"""按标题/段落分块"""
|
||||
chunks: list[str] = []
|
||||
cur: list[str] = []
|
||||
for line in text.splitlines():
|
||||
if line.startswith("#"):
|
||||
if cur:
|
||||
chunks.append("\n".join(cur).strip())
|
||||
cur = [line]
|
||||
else:
|
||||
cur.append(line)
|
||||
if cur:
|
||||
chunks.append("\n".join(cur).strip())
|
||||
return [c for c in chunks if len(c) > 10]
|
||||
|
||||
|
||||
def _embed(texts: list[str]) -> list[list[float]]:
|
||||
"""dashscope text-embedding-v3 批量向量化"""
|
||||
import dashscope
|
||||
|
||||
api_key = os.environ.get("DASHSCOPE_API_KEY") or ""
|
||||
resp = dashscope.TextEmbedding.call(
|
||||
model=EMBED_MODEL,
|
||||
input=texts,
|
||||
api_key=api_key,
|
||||
)
|
||||
if resp.status_code != 200:
|
||||
raise RuntimeError(f"embedding API {resp.status_code}: {resp.code} {resp.message}")
|
||||
out = [e["embedding"] for e in resp.output["embeddings"]]
|
||||
return out
|
||||
|
||||
|
||||
def _load_index():
|
||||
"""懒加载索引；缺失时构建并缓存"""
|
||||
global _index
|
||||
if _index is not None:
|
||||
return _index
|
||||
if INDEX_FILE.exists():
|
||||
try:
|
||||
_index = json.loads(INDEX_FILE.read_text("utf-8"))
|
||||
log.info("rag: 知识索引已加载（%d 块）", len(_index["chunks"]))
|
||||
return _index
|
||||
except Exception: # noqa: BLE001
|
||||
log.warning("rag: 索引损坏，重建")
|
||||
# 构建
|
||||
chunks = _chunk_md(KB_FILE.read_text("utf-8"))
|
||||
log.info("rag: 构建知识索引（%d 块，embedding 中…）", len(chunks))
|
||||
embs = _embed(chunks)
|
||||
_index = {"chunks": chunks, "embeddings": embs}
|
||||
INDEX_FILE.write_text(json.dumps(_index, ensure_ascii=False), "utf-8")
|
||||
log.info("rag: 知识索引构建完成 -> %s", INDEX_FILE)
|
||||
return _index
|
||||
|
||||
|
||||
def retrieve(query: str, top_k: int = 3) -> list[str]:
|
||||
"""检索与 query 最相关的知识段落"""
|
||||
if not query or not query.strip():
|
||||
return []
|
||||
idx = _load_index()
|
||||
q = _embed([query])[0]
|
||||
sims = np.dot(np.array(idx["embeddings"]), np.array(q))
|
||||
top = np.argsort(-sims)[:top_k]
|
||||
return [idx["chunks"][int(i)] for i in top if float(sims[int(i)]) > 0.3]
|
||||
|
||||
|
||||
def brief(max_chars: int = 1200) -> str:
|
||||
"""固定知识摘要（前端拼进 instructions，静态知识一次注入）"""
|
||||
idx = _load_index()
|
||||
out = []
|
||||
total = 0
|
||||
for c in idx["chunks"]:
|
||||
if total + len(c) > max_chars:
|
||||
break
|
||||
out.append(c)
|
||||
total += len(c)
|
||||
return "\n\n".join(out)
|
||||
@@ -426,3 +426,42 @@ def vision_frame(body: VisionFrameBody):
|
||||
return {"ok": False, "error": "missing image"}
|
||||
from .vision_yolo import predict_base64
|
||||
return predict_base64(img_b64)
|
||||
|
||||
|
||||
# ==================== 智能体工具调用 + 知识库 ====================
|
||||
class ToolsExecBody(BaseModel):
|
||||
name: str = ""
|
||||
args: dict = {}
|
||||
|
||||
@router.post("/api/tools/exec")
|
||||
def tools_exec(body: ToolsExecBody):
|
||||
"""执行智能体工具（LLM function calling）：name + args → 结果字符串"""
|
||||
name = (body.name or "").strip()
|
||||
if not name:
|
||||
return {"ok": False, "error": "missing tool name"}
|
||||
from .tools import exec_tool
|
||||
result = exec_tool(name, body.args or {})
|
||||
return {"ok": True, "result": result}
|
||||
|
||||
@router.get("/api/kb/brief")
|
||||
def kb_brief(max_chars: int = 1200):
|
||||
"""园区知识摘要（前端拼进对话 instructions）"""
|
||||
try:
|
||||
from .rag import brief
|
||||
return {"ok": True, "brief": brief(max_chars=max_chars)}
|
||||
except Exception as e: # noqa: BLE001
|
||||
log.warning("kb brief 失败: %s", e)
|
||||
return {"ok": False, "error": str(e)}
|
||||
|
||||
@router.get("/api/kb/retrieve")
|
||||
def kb_retrieve(q: str = "", top_k: int = 3):
|
||||
"""知识库检索（query → 相关段落）"""
|
||||
if not q.strip():
|
||||
return {"ok": False, "error": "missing q"}
|
||||
try:
|
||||
from .rag import retrieve
|
||||
chunks = retrieve(q, top_k=top_k)
|
||||
return {"ok": True, "chunks": chunks}
|
||||
except Exception as e: # noqa: BLE001
|
||||
log.warning("kb retrieve 失败: %s", e)
|
||||
return {"ok": False, "error": str(e)}
|
||||
|
||||
@@ -0,0 +1,127 @@
|
||||
# -*- coding: utf-8 -*-
|
||||
"""智能体工具注册表 —— LLM function calling 的后端执行器
|
||||
|
||||
- TOOLS：工具名 → 描述 + JSON Schema 参数（给 LLM 声明用）
|
||||
- exec_tool(name, args)：执行工具，返回给 LLM 的字符串结果（JSON）
|
||||
- 供 POST /api/tools/exec 调用；前端 s2s 客户端收到 toolcall 事件后转发到此执行
|
||||
"""
|
||||
import json
|
||||
import logging
|
||||
from datetime import datetime
|
||||
|
||||
log = logging.getLogger("dpm.tools")
|
||||
|
||||
|
||||
# ── 工具实现 ────────────────────────────────────────────────────────────
|
||||
|
||||
def _get_park_overview(args):
|
||||
"""园区实时概览（在园项目/就业/营收/在园人数等）"""
|
||||
from .sim_engine import sim_engine
|
||||
d = sim_engine.snapshot()
|
||||
return json.dumps({
|
||||
"在园项目": d["projects"]["inPark"],
|
||||
"累计孵化企业": d["projects"]["cum"],
|
||||
"今日新增项目": d["projects"]["todayNew"],
|
||||
"带动就业(人)": d["jobs"]["total"],
|
||||
"今日新增就业": d["jobs"]["todayNew"],
|
||||
"累计营收(万元)": d["revenue"]["total"],
|
||||
"今日营收(万元)": d["revenue"]["today"],
|
||||
"营收同比增速(%)": d["revenue"]["growth"],
|
||||
"设备在线率(%)": d["park"]["devices"],
|
||||
"今日能耗(kWh)": d["park"]["energy"],
|
||||
"当前在园人数": d["park"]["people"],
|
||||
"会议预约": d["park"]["meeting"],
|
||||
"工位使用": d["park"]["desk"],
|
||||
}, ensure_ascii=False)
|
||||
|
||||
|
||||
def _query_companies(args):
|
||||
"""园区入驻企业名录查询（关键词过滤）"""
|
||||
from .sim_engine import COMPANY_NAMES
|
||||
q = (args.get("keyword") or "").strip()
|
||||
names = [n for n in COMPANY_NAMES if q in n] if q else list(COMPANY_NAMES)
|
||||
return json.dumps({
|
||||
"total": len(names),
|
||||
"keyword": q,
|
||||
"companies": names[:20],
|
||||
}, ensure_ascii=False)
|
||||
|
||||
|
||||
def _control_display(args):
|
||||
"""大屏显示控制：切页 / 弹通知 / 媒体播放暂停（经 MQTT 广播到大屏）"""
|
||||
from .mqtt import hub
|
||||
action = (args.get("action") or "").strip()
|
||||
target = (args.get("target") or "").strip()
|
||||
if action not in ("switch_page", "alert", "media_play", "media_pause"):
|
||||
return json.dumps({"error": f"不支持的 action: {action}"}, ensure_ascii=False)
|
||||
params = {"target": target} if target else {}
|
||||
ok = hub.publish_command(action, params).get("published", False)
|
||||
return json.dumps({"ok": ok, "action": action, "target": target}, ensure_ascii=False)
|
||||
|
||||
|
||||
def _get_time(args):
|
||||
"""当前日期时间"""
|
||||
return json.dumps({"datetime": datetime.now().strftime("%Y-%m-%d %H:%M:%S")}, ensure_ascii=False)
|
||||
|
||||
|
||||
# ── 注册表 ──────────────────────────────────────────────────────────────
|
||||
|
||||
TOOLS = {
|
||||
"get_park_overview": {
|
||||
"fn": _get_park_overview,
|
||||
"description": "获取园区实时运营概览：在园项目数、累计孵化企业、带动就业、营收（今日/累计）、设备在线率、在园人数、能耗等。回答园区数据类问题时使用。",
|
||||
"parameters": {"type": "object", "properties": {}},
|
||||
},
|
||||
"query_companies": {
|
||||
"fn": _query_companies,
|
||||
"description": "查询园区入驻企业名录，支持按企业名关键词过滤。回答'有哪些企业/某企业是否入驻'时使用。",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {"keyword": {"type": "string", "description": "企业名关键词，可为空字符串"}},
|
||||
},
|
||||
},
|
||||
"control_display": {
|
||||
"fn": _control_display,
|
||||
"description": "控制大屏显示：switch_page 切换页面、alert 弹通知、media_play/media_pause 控制媒体播放。用户要求控制大屏时使用。",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"action": {"type": "string", "enum": ["switch_page", "alert", "media_play", "media_pause"]},
|
||||
"target": {"type": "string", "description": "目标：页面路径（如 /、/twin、/ai、/voice）或通知文本"},
|
||||
},
|
||||
},
|
||||
},
|
||||
"get_time": {
|
||||
"fn": _get_time,
|
||||
"description": "获取当前日期时间。用户问'现在几点/今天几号'时使用。",
|
||||
"parameters": {"type": "object", "properties": {}},
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
def tool_schemas():
|
||||
"""给 LLM 的 tools 定义（OpenAI function calling 格式）"""
|
||||
return [
|
||||
{
|
||||
"type": "function",
|
||||
"name": name,
|
||||
"description": t["description"],
|
||||
"parameters": t["parameters"],
|
||||
}
|
||||
for name, t in TOOLS.items()
|
||||
]
|
||||
|
||||
|
||||
def exec_tool(name: str, args: dict):
|
||||
"""执行工具 → 返回给 LLM 的字符串结果（JSON）"""
|
||||
t = TOOLS.get(name)
|
||||
if not t:
|
||||
log.warning("tools: 未知工具 %s", name)
|
||||
return json.dumps({"error": f"未知工具: {name}"}, ensure_ascii=False)
|
||||
try:
|
||||
result = t["fn"](args or {})
|
||||
log.info("tools: %s args=%s -> %s", name, json.dumps(args, ensure_ascii=False)[:120], str(result)[:120])
|
||||
return result
|
||||
except Exception as e: # noqa: BLE001
|
||||
log.error("tools: %s 执行失败 %s", name, e, exc_info=True)
|
||||
return json.dumps({"error": f"{name} 执行失败: {e}"}, ensure_ascii=False)
|
||||
Reference in New Issue
Block a user