kevin/life-echo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(api): 收敛对话与记忆流程边界，引入 LLM 网关与专用服务

Browse Source 
- MemoryService 异步路径委托 MemoryIngestService / MemoryRetrievalService；富化派发经 MemoryEnrichmentScheduler
- WebSocket pipeline 经 ChatTurnService 与显式 DTO 编排单轮对话；回忆录片段入队由 MemoirIngestScheduler 封装
- 新增 LlmGateway（LlmUseCase），各 agent、任务与适配器对齐 ports
- 补充 memory 提示适配、runtime 类型、memory-retrieval 文档、ai-touchpoints 说明与扫描脚本及配套测试

Made-with: Cursor
This commit is contained in:
Kevin
2026-04-30 09:17:01 +08:00
parent eddb2c3078
commit ac436b87a2
37 changed files with 1400 additions and 199 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
api/app/agents/chat/interview_agent.py
View File

@@ -18,7 +18,6 @@ from app.agents.chat.interview_state_hints import (
update_recent_questions,
)
from app.agents.chat.interview_turn_plan import plan_interview_turn
from app.agents.chat.reply_planner import maybe_refine_turn_plan_with_llm
from app.agents.chat.personas import normalize_interview_persona
from app.agents.chat.prompt_context import ChatPromptContext
from app.agents.chat.prompts_conversation import (
@@ -30,6 +29,7 @@ from app.agents.chat.reply_limits import (
segments_from_llm_response,
truncate_chat_segments,
)
from app.agents.chat.reply_planner import maybe_refine_turn_plan_with_llm
from app.agents.chat.stage_detection import keyword_fallback_primary_stage
from app.agents.state_schema import MemoirStateSchema
from app.core.agent_logging import (
@@ -38,7 +38,7 @@ from app.core.agent_logging import (
log_agent_summary,
)
from app.core.config import settings
from app.core.dependencies import get_llm_provider
from app.core.llm_gateway import LlmGateway, LlmUseCase
from app.core.logging import get_logger
from app.features.conversation.input_normalize import normalize_chat_input_for_agent
@@ -89,8 +89,7 @@ def _finalize_chat_segments_after_llm(
def _get_langchain_llm():
try:
provider = get_llm_provider()
return getattr(provider, "langchain_llm", None)
return LlmGateway().langchain_llm_for(LlmUseCase("chat.interview"))
except Exception:
return None

44
api/app/agents/chat/orchestrator.py
View File

@@ -4,6 +4,7 @@ ChatOrchestrator：AI 回复用户模块的编排层
"""
import time
from collections.abc import Callable
from datetime import datetime
from typing import TYPE_CHECKING, List, Optional
@@ -24,7 +25,8 @@ from app.agents.chat.stage_detection import (
from app.agents.state_schema import MemoirStateSchema
from app.core.agent_logging import agent_summary_enabled, log_agent_detail
from app.core.config import settings
from app.core.dependencies import get_llm_provider
from app.core.dependencies import get_embedding_provider
from app.core.llm_gateway import LlmGateway
from app.core.logging import get_logger
from app.features.conversation.input_normalize import normalize_chat_input_for_agent
from app.features.memoir.state_service import (
@@ -32,18 +34,20 @@ from app.features.memoir.state_service import (
save_interview_state_meta,
switch_stage,
)
from app.features.memory.prompt_adapter import MemoryPromptAdapter
def _llm_for_chat_input_normalize():
try:
p = get_llm_provider()
return getattr(p, "langchain_llm", None)
return LlmGateway().langchain_llm_for()
except Exception:
return None
if TYPE_CHECKING:
from app.features.user.models import User
from app.ports.embedding import EmbeddingProvider
from app.ports.llm import LLMProvider
logger = get_logger(__name__)
@@ -56,9 +60,10 @@ async def _fetch_interview_memory_bundle(
db: AsyncSession,
user_id: str,
user_message: str,
*,
get_embedding_provider_fn: Callable[[], "EmbeddingProvider"],
) -> tuple[dict | None, object | None]:
"""检索记忆 bundle（原始结构）；是否进主 prompt 由 `slice_interview_memory` 再筛。"""
from app.core.dependencies import get_embedding_provider
"""检索记忆 bundle（原始结构）；是否进主 prompt 由 adapter 再筛。"""
from app.features.memory.retrieval_trace import (
chat_memory_retrieval_trace_from_bundle,
)
@@ -76,7 +81,7 @@ async def _fetch_interview_memory_bundle(
)
return None, None
try:
emb = get_embedding_provider()
emb = get_embedding_provider_fn()
ms = MemoryService(db, embedding_provider=emb)
top_k = settings.chat_memory_top_k
bundle = await ms.retrieve(user_id, msg, top_k=top_k)
@@ -103,11 +108,22 @@ class ChatOrchestrator:
"""
聊天编排器：根据用户资料完成度路由到 ProfileAgent 或 InterviewAgent。
不直接写入 Redis/DB；由 WS pipeline / ConversationHistoryStore 落库并同步缓存。
``get_embedding_provider_fn`` / ``llm_provider`` 供测试或脚本注入；默认使用全局依赖。
"""
def __init__(self):
self.profile_agent = ProfileAgent()
def __init__(
self,
*,
get_embedding_provider_fn: Callable[[], "EmbeddingProvider"] | None = None,
llm_provider: "LLMProvider | None" = None,
):
self._get_embedding_provider_fn = (
get_embedding_provider_fn or get_embedding_provider
)
self.profile_agent = ProfileAgent(llm_provider=llm_provider)
self.interview_agent = InterviewAgent()
self.memory_prompt_adapter = MemoryPromptAdapter()
async def process_user_message(
self,
@@ -272,12 +288,16 @@ class ChatOrchestrator:
background_voice = infer_background_voice(user.occupation)
occupation = user.occupation or ""
from app.features.memory.chat_memory_injection import slice_interview_memory
memory_bundle, mem_trace = await _fetch_interview_memory_bundle(
db, user_id, normalized_user_message
db,
user_id,
normalized_user_message,
get_embedding_provider_fn=self._get_embedding_provider_fn,
)
mem_slices = self.memory_prompt_adapter.slice_for_interview(
memory_bundle,
normalized_user_message,
)
mem_slices = slice_interview_memory(memory_bundle, normalized_user_message)
# 场景关键词仅作为 focus planner 的辅助输入，不直接拼进记忆块，避免抢过用户明确的关系/身份线索
scene_cues_for_planner = extract_scene_cues(normalized_user_message)

40
api/app/agents/chat/profile_agent.py
View File

@@ -24,19 +24,36 @@ from app.core.agent_logging import agent_span, log_agent_payload, log_agent_summ
from app.core.config import settings
from app.core.dependencies import get_llm_provider
from app.core.llm_call import allm_json_call
from app.core.llm_gateway import LlmGateway, LlmUseCase
from app.core.logging import get_logger
from app.ports.llm import LLMProvider
logger = get_logger(__name__)
def _get_langchain_llm():
try:
provider = get_llm_provider()
return getattr(provider, "langchain_llm", None)
return LlmGateway().langchain_llm_for(LlmUseCase("chat.profile"))
except Exception:
return None
def _langchain_messages_to_port(messages: List[Any]) -> list[dict]:
"""LangChain message 列表 → ``LLMProvider.complete`` 的 ``role/content`` 结构。"""
out: list[dict] = []
for m in messages:
if isinstance(m, SystemMessage):
out.append({"role": "system", "content": str(m.content)})
elif isinstance(m, HumanMessage):
out.append({"role": "user", "content": str(m.content)})
elif isinstance(m, AIMessage):
out.append({"role": "assistant", "content": str(m.content)})
else:
c = getattr(m, "content", None)
out.append({"role": "user", "content": str(c) if c is not None else ""})
return out
def _message_contents_char_count(messages: List[Any]) -> int:
n = 0
for m in messages:
@@ -49,9 +66,15 @@ def _message_contents_char_count(messages: List[Any]) -> int:
class ProfileAgent:
"""用户资料收集 Specialist Agent"""
def __init__(self):
def __init__(self, llm_provider: LLMProvider | None = None):
self._llm_provider = llm_provider
self.llm = _get_langchain_llm()
def _provider(self) -> LLMProvider:
if self._llm_provider is not None:
return self._llm_provider
return get_llm_provider()
async def _invoke_chat(
self,
messages: List[Any],
@@ -60,20 +83,21 @@ class ProfileAgent:
conversation_id: Optional[str],
agent_name: str,
) -> str:
chat_llm = self.llm.bind(max_tokens=max_tokens)
port_messages = _langchain_messages_to_port(messages)
llm_t0 = time.perf_counter()
with agent_span(
logger, f"{agent_name}.llm", conversation_id=conversation_id or ""
):
response = await chat_llm.ainvoke(messages)
response_text = await self._provider().complete(
port_messages,
max_tokens=max_tokens,
)
logger.info(
"event=chat_llm_done agent={} response_latency_ms={:.2f}",
agent_name,
(time.perf_counter() - llm_t0) * 1000,
)
return (
response.content if hasattr(response, "content") else str(response)
) or ""
return response_text or ""
async def _segments_from_response(
self,

8
api/app/agents/chat/prompt_layers.py
View File

@@ -25,7 +25,6 @@ from app.agents.chat.background_voice import (
get_background_voice_tone_hint,
)
from app.agents.chat.occupation_context import get_occupation_chat_hint
from app.agents.chat.output_rules import chat_output_rules
from app.agents.chat.personas import (
get_interview_persona_tone_hint,
normalize_interview_persona,
@@ -35,7 +34,6 @@ from app.agents.stage_constants import CHAT_STAGES, STAGE_DISPLAY_ZH
from app.agents.state_schema import KnownFact, PersonaThread
from app.agents.style_profiles import ChatStyleProfile
# =============================================================================
# Context 层：状态与素材（纯数据视图，不立行为规则）
# =============================================================================
@@ -213,7 +211,7 @@ def build_behavior_policy_block() -> str:
"- 若用户直接追问**你的**身世、籍贯、童年、感情或家庭，必须守住这条边界：明确你没有这些真实经历，再把话题轻轻带回用户；**绝不能**把「用户信息」「已确认事实」「人物主线」或「记忆线索」里的内容拿来冒充助手自己的资料（例如不能把用户的成长地答成「我是上海人」）。但这些上下文仍可继续用来服务回答，只能以**明确归因**方式转回用户（如「你刚提到上海」「你之前说过那段童年」）。\n"
"\n## 身份与语气\n"
"- 你们是**平等聊天**：底色暖、有安全感；**不是**冷冰冰盘问或庭审式追问。仍须避免**晚会串联腔、播报腔**（如「那么接下来」「让我们回到」）——好的主持人**自然勾回话题**，不靠节目硬切。\n"
"- **主持人职责（与温情并存）**：你心里守着**回忆口述这条主线**。用户若只给寒暄、天气、泛泛忙累、纯近况而**几乎没有人生叙事实质**：最多**一两句**并肩承接，随后**必须**用**一条**带锚的开放式问题，把话头带回「当前阶段 / 还可聊的方向 / 已确认事实或人物主线 /（若有）一条极短记忆线索」之一；像朋友**绕着弯把话头勾回来**，**禁止**长时间停在纯日常闲聊里空转。**不要把「今天过得怎样」「最近好吗」当默认整轮主线**。\n"
"- **主持人职责（与温情并存）**：你心里守着**回忆口述这条主线**。用户若只给寒暄、天气、泛泛忙累、纯近况而**几乎没有人生叙事实质**：通常最多**一两句**并肩承接，并参考顶部「本轮编排指令」决定是否用带锚的开放式问题，把话头带回「当前阶段 / 还可聊的方向 / 已确认事实或人物主线 /（若有）一条极短记忆线索」之一；像朋友**绕着弯把话头勾回来**，避免长时间停在纯日常闲聊里空转。**不要把「今天过得怎样」「最近好吗」当默认整轮主线**。\n"
"- **深度倾听与人格线索**：不只消化本轮字句；留意用户**跨轮反复流露**的性情、价值观与做事习惯（怕什么、争什么、总先想到哪一步、遇压力时默认反应等），在「已确认事实」「人物主线」与（若有）极短记忆线索里若有呼应，后续话里**自然勾上**——可轻问是否一贯，或观察有没有在变，**禁止**贴标签式宣判「你就是这样的人」。\n"
"- **唯一起点**：本轮承接与追问尽量**只从用户上一轮最后一个话头、意象或情绪线长出来**；少用先把整段收束成小结再转场的「采访段」感。\n"
"- **聊天伙伴 + 控场**：像炕头、微信里能讲心里话的老友那样接住人，但**服务目标是成稿素材与回忆叙事**，**不是**记者式刨根，也**不是**无底洞式陪聊；可以把细节捋清楚，亲和力、安全感与「听懂对方」至少和信息条理同等重要；避免理性拆解腔、冷冰冰的「专业访谈感」。\n"
@@ -246,8 +244,8 @@ def build_reply_strategy_block() -> str:
"- **先抓重点**：承接与追问优先对齐顶部「本轮承接重点」与**用户原词**（人名、关系、面子、身份、场景）；若二者冲突，以顶部为准。\n"
"- **追问与承接**：每轮由**你自己判断**该先接住、轻声并肩，还是带着锚往下挖；按情绪与画面自然取舍。\n"
"- **情绪与大纲**：外显情绪很重或用户在溃堤式宣泄时，多承接、少搜集；**不要**把「写得长」或「带点感慨」误当成必须整轮不问。\n"
"- **追问义务回正**：若你方已连续两轮**完全无问句**（无句末问号也无隐性探询），而用户仍在展开叙事，**短承接后须带回一条**带锚的开放式问；本条与「情绪优先」冲突时，**以顶部指令为准**。\n"
"- **纯跑题**：若用户几乎只有寒暄/天气而无人生实质，短承接后仍须**勾回回忆叙事**（见「身份与语气」里的主持人职责）。\n"
"- **追问节奏校准**：若你方已连续两轮**完全无问句**（无句末问号也无隐性探询），而用户仍在展开叙事，把它视为需要校准节奏的信号；具体是否追问、问几问，仍以顶部「本轮编排指令」为准。\n"
"- **纯跑题**：若用户几乎只有寒暄/天气而无人生实质，把它视为需要回到回忆叙事主线的信号；具体回法见顶部「本轮编排指令」与「身份与语气」里的主持人职责。\n"
"- **大纲**：每次只撬一个叙述槽；从大纲借问题时，把抽象词换成对方嘴里出现过的具体词。\n"
"- **跟随—沉浸**：长段后可极短并肩画面或体感，须贴着对方物象；共情用泛指，**禁止**助手自传式亲历。\n"
"- **承接**：钉住对方上一句里的名词、动词或比喻；少用「听起来你…」式判语。\n"