zhanghongyu_0626
7dadec279d
- Implemented `_merged_items` and `_merged_memory_text` functions to consolidate consult and chitchat memories, improving context awareness in intent classification and response generation. - Updated intent prompts to include recent dialogue history, aiding in the resolution of ambiguous user queries. - Enhanced `search_knowledge` tool to utilize context window for better query understanding, addressing issues with omitted references in user inputs. - Fixed existing test cases to reflect changes in intent constants and ensure accurate context handling during tests. This update significantly improves the handling of multi-turn dialogues, ensuring a more coherent and contextually aware interaction for users.
78 lines
3.0 KiB
Python
78 lines
3.0 KiB
Python
"""知识库对话 Tool（T21-5 · FLOW §2「milvus_tool：产品规则 RAG + source_refs」）。
|
||
|
||
与 core_tools / risk.chat_tools 同构的 Tool 定义层：由 tool_service.run_tool
|
||
经统一注册表分发。search_knowledge 是公开知识检索（产品手册/交易规则）——
|
||
**不做客户归属校验**（ToolSpec.skip_access_check=True）：
|
||
|
||
- 知识内容为公开产品信息，无客户数据，归属校验无对象；
|
||
- 开放范围（customer + advisor，用户拍板 2026-09-07）由意图层控制：
|
||
_INTENT_KEYWORDS 仅 customer/advisor 组配置 kb 关键词，risk/analyst
|
||
不命中 → Tool 不会被触发；即使被直接调用，返回的也只是公开知识。
|
||
|
||
func 签名对齐 runner 约定 ``func(customer_id, core_ro, risk_repo, **params)``；
|
||
query 为唯一业务参数（tool_node 从 user_message 构造 tool_input 注入，
|
||
一期不来自 LLM 输出）。RAG 异常（EmbeddingError 等）原样上抛 → run_tool
|
||
统一转 TOOL_ERROR 留痕，不静默降级空结果（防 LLM 编造回答）。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
from typing import Any, Callable
|
||
|
||
from app.service import rag_service
|
||
|
||
# 对话 Tool 检索条数（与 rag_service.DEFAULT_TOP_K 一致；独立常量便于对话口径单独调整）
|
||
KB_TOP_K = 3
|
||
|
||
|
||
def search_knowledge(
|
||
query: str = "",
|
||
customer_id: str = "",
|
||
core_ro=None,
|
||
risk_repo=None,
|
||
context_window: str = "",
|
||
) -> dict[str, Any]:
|
||
"""产品知识检索（TopK chunks + 溯源清单）。
|
||
|
||
customer_id/core_ro/risk_repo 为 runner 恒传参数，本 Tool 不使用
|
||
（公开知识，无归属语义）；保留形参以满足统一签名。
|
||
context_window 为 tool_node 注入的近期对话，拼接进 query 使检索
|
||
感知历史（解决「那申购呢」类省略指代）。
|
||
"""
|
||
effective_query = query
|
||
if context_window and context_window.strip():
|
||
effective_query = f"{context_window.strip()}\n{query}"
|
||
out = rag_service.search_knowledge(effective_query, top_k=KB_TOP_K)
|
||
return {
|
||
"hit_count": len(out["results"]),
|
||
"results": out["results"],
|
||
"source_refs": out["source_refs"],
|
||
}
|
||
|
||
|
||
class KBToolSpec(dict):
|
||
"""知识库 Tool 注册表条目（与 core_tools.ToolSpec / RiskToolSpec 同构；
|
||
新增 skip_access_check：True → run_tool 跳过客户归属校验）。"""
|
||
|
||
|
||
KB_TOOL_REGISTRY: dict[str, KBToolSpec] = {
|
||
"search_knowledge": KBToolSpec(
|
||
func=search_knowledge,
|
||
description="检索产品知识库（基金产品手册、费率、申赎规则、风险说明；返回结果附溯源）",
|
||
requires_customer=False,
|
||
param_whitelist=("query",),
|
||
int_bounds={},
|
||
skip_access_check=True,
|
||
),
|
||
}
|
||
|
||
|
||
def get_kb_tool(name: str) -> KBToolSpec | None:
|
||
"""白名单查找（未知 Tool 返回 None）。"""
|
||
return KB_TOOL_REGISTRY.get(name)
|
||
|
||
|
||
def kb_tool_func(name: str) -> Callable[..., dict[str, Any]] | None:
|
||
spec = KB_TOOL_REGISTRY.get(name)
|
||
return spec["func"] if spec else None
|