59d4b19d7d
- Phase1/2：移除 MemoirOrchestrator.run 与 process_memoir_segments 别名；文档改为 process_memoir_phase1。 - 槽位校验集中到 stage_constants（filter_stage_slots），批处理与顺序路径及 state_service 写库一致。 - StoryRoute：no_llm/parse_error/invalid_target 保守 new_story；短篇护栏不覆盖这些 fallback。 - Phase2 低置信单路径可选延迟（StoryPipelineResult.deferred）：不写 Chapter/Story，Segment 记录 defer 元数据，冷却内不重复消费；上限后停自动重试，Phase1 同类目新段唤醒池内段。 - Alembic 0017：segments 表 narrative_defer_* 列。 - ProfileAgent：经 LlmGateway/注入 Provider 统一聊天与 JSON，新增测试。 - ImagePromptOrchestrator：LLM 初始化失败可依配置降级或硬失败；补充策略测试。 - 配套单测与 README/本地开发文档表述更新。 Co-authored-by: Cursor <cursoragent@cursor.com>
473 lines
27 KiB
Python
473 lines
27 KiB
Python
"""
|
||
统一配置：所有环境变量通过此模块的 Settings 单点读取。
|
||
业务代码只允许 import settings，禁止散落 os.getenv() / load_dotenv()。
|
||
|
||
本地开发时由 api/development.sh 在启动前将 .env.development 同步为 .env（每次启动覆盖）。
|
||
Docker / 服务端由镜像与 compose 注入进程环境；此处仅固定读取工作目录下的 .env 作为默认值来源。
|
||
进程环境变量（容器 environment、export）覆盖 .env 同名项。
|
||
"""
|
||
|
||
import secrets
|
||
|
||
from pydantic import AliasChoices, Field, field_validator
|
||
from pydantic_settings import BaseSettings, SettingsConfigDict
|
||
|
||
|
||
class Settings(BaseSettings):
|
||
model_config = SettingsConfigDict(
|
||
env_file=".env",
|
||
env_file_encoding="utf-8",
|
||
case_sensitive=False,
|
||
extra="ignore",
|
||
)
|
||
|
||
# ── Database ──────────────────────────────────────────────
|
||
database_url: str = "postgresql://postgres:postgres@localhost:5432/life_echo"
|
||
# 启动时是否执行 Alembic（main.py lifespan）；测试或仅读场景可关
|
||
alembic_run_on_startup: bool = True
|
||
# True：迁移失败则进程退出（生产推荐）。False：仅打错误日志并继续（本地无 DB 时）
|
||
alembic_startup_fail_fast: bool = False
|
||
alembic_startup_max_retries: int = Field(default=3, ge=1, le=10)
|
||
alembic_startup_retry_base_seconds: float = Field(default=1.0, ge=0.1, le=60.0)
|
||
|
||
# ── Redis ─────────────────────────────────────────────────
|
||
redis_url: str = "redis://localhost:6379/0"
|
||
redis_session_ttl: int = 86400
|
||
|
||
# ── Runtime / Celery 开发体验 ─────────────────────────────
|
||
# APP_ENV：本地默认 development；Docker 生产栈请设为 production
|
||
app_environment: str = Field(
|
||
default="development",
|
||
validation_alias=AliasChoices("APP_ENV", "APP_ENVIRONMENT"),
|
||
)
|
||
# 非 production 且为 True 时，在 main/internal_main 连接 Redis 后清空 Celery 队列（不 FLUSHDB，不影响会话键）
|
||
celery_purge_broker_on_startup: bool = False
|
||
# Memory LLM 富化任务路由队列；可与主 worker 分离（见 README / docker-compose）
|
||
celery_memory_enrichment_queue: str = "memory_idle"
|
||
|
||
# ── Auth / JWT ────────────────────────────────────────────
|
||
secret_key: str = Field(default_factory=lambda: secrets.token_urlsafe(32))
|
||
algorithm: str = "HS256"
|
||
access_token_expire_minutes: int = 120
|
||
refresh_token_expire_days: int = 30
|
||
# 本地/内网评测：允许 POST /api/auth/mock/sms-login 跳过短信（须显式开启；production 下路由仍拒绝）
|
||
mock_sms_login_enabled: bool = False
|
||
|
||
# ── LLM / DeepSeek ───────────────────────────────────────
|
||
deepseek_api_key: str = ""
|
||
deepseek_base_url: str = "https://api.deepseek.com"
|
||
# 官方新模型名（V4-Flash）；与弃用名 deepseek-chat 对齐为「非思考」需另设 deepseek_thinking_enabled
|
||
deepseek_model: str = "deepseek-v4-flash"
|
||
# V4-Flash 在官方 API 中 thinking 默认为 enabled；主链路为对齐旧版 deepseek-chat 默认关闭
|
||
deepseek_thinking_enabled: bool = False
|
||
llm_api_key: str = ""
|
||
llm_base_url: str = ""
|
||
llm_model: str = ""
|
||
llm_temperature: float = 0.7
|
||
# 空字符串：快档位与默认模型相同；分类/抽取/记忆富化等可单独指定较轻模型
|
||
llm_fast_model: str = ""
|
||
|
||
# ── Memory 向量（智谱 BigModel 国内 embedding-3；与 LLM/DeepSeek 密钥分离）──
|
||
zhipu_api_key: str = ""
|
||
embedding_base_url: str = "https://open.bigmodel.cn/api/paas/v4"
|
||
embedding_model: str = "embedding-3"
|
||
|
||
# ── Chat 访谈（token 上限 + 代码截断，见 reply_limits）──
|
||
chat_interview_max_tokens: int = 512
|
||
chat_interview_max_segments: int = 2
|
||
chat_interview_max_chars_per_segment: int = 380
|
||
chat_opening_max_tokens: int = 380
|
||
chat_profile_followup_max_tokens: int = 280
|
||
# Redis 全量历史仅用于 turn 计数；注入 LLM 时截取最近若干轮与字符预算
|
||
chat_history_max_pairs: int = Field(default=15, ge=1, le=500)
|
||
chat_history_max_chars: int = Field(default=6000, ge=256, le=500_000)
|
||
chat_era_context_enabled: bool = True
|
||
# 访谈：每轮用 LLM 判定用户主人生阶段并更新 MemoirState.current_stage；False 时仅用关键词
|
||
chat_stage_detection_enabled: bool = True
|
||
chat_stage_detection_max_tokens: int = 128
|
||
# 访谈性格：default | warm_listener | curious_guide（未知值按 default）
|
||
chat_interview_persona: str = "default"
|
||
# 访谈/开场 LLM 采样温度：略高于通用 llm_temperature，利于口语与叙事变化、减程式句
|
||
chat_interview_temperature: float = Field(default=0.93, ge=0.0, le=2.0)
|
||
# 访谈：按用户本轮话检索记忆并注入 prompt（关则不调 MemoryService.retrieve）
|
||
chat_memory_retrieval_enabled: bool = True
|
||
chat_memory_top_k: int = Field(default=8, ge=1, le=30)
|
||
chat_memory_evidence_max_chars: int = Field(default=4096, ge=256, le=50_000)
|
||
# 访谈记忆注入使用聊天专用安全格式化（编号引用 + 主语弱化说明）
|
||
chat_memory_safe_evidence_format_enabled: bool = True
|
||
# True：在规则 TurnPlan 之后追加一轮轻量 JSON focus planner（本轮承接重点 + memory 引用 + 回复形状；失败则回退基线）
|
||
chat_reply_planner_llm_enabled: bool = False
|
||
chat_reply_planner_max_tokens: int = Field(default=256, ge=64, le=1024)
|
||
chat_reply_planner_temperature: float = Field(default=0.2, ge=0.0, le=1.0)
|
||
|
||
# ── Memoir 叙事忠实度检查（FidelityCheckAgent）────────────────
|
||
memoir_fidelity_check_enabled: bool = True
|
||
memoir_fidelity_check_max_tokens: int = 512
|
||
# 口述归一（进入叙事 / 忠实度前；segment 原文不落库）：off | rules | llm
|
||
memoir_oral_normalize_enabled: bool = True
|
||
memoir_oral_normalize_mode: str = "rules"
|
||
memoir_oral_normalize_llm_max_tokens: int = Field(default=512, ge=64, le=4096)
|
||
memoir_oral_normalize_llm_max_input_chars: int = Field(
|
||
default=8000, ge=64, le=50_000
|
||
)
|
||
# 聊天：模型消费净稿（不改变 segment 落库原文）；与 memoir 规则层共用，配置独立
|
||
chat_input_normalize_enabled: bool = True
|
||
chat_input_normalize_mode: str = "rules" # off | rules | llm
|
||
chat_input_normalize_llm_max_tokens: int = Field(default=512, ge=64, le=4096)
|
||
chat_input_normalize_llm_max_input_chars: int = Field(
|
||
default=8000, ge=64, le=50_000
|
||
)
|
||
# True 且 mode=llm：仅语音/ASR 段走 LLM 纠错；键盘输入仅规则归一（省每轮 LLM）
|
||
chat_input_normalize_llm_voice_only: bool = True
|
||
# 资料收集：超过该对话轮次（Redis 全量轮次计数）仍有缺失字段时，强制进入访谈，避免长期问卷感
|
||
chat_profile_max_turns: int = Field(default=8, ge=1, le=500)
|
||
|
||
# Memoir Phase1：多 segment 一批一次 LLM 完成抽取+章节分类（失败回退逐段）；单段且关时仍逐段
|
||
memoir_phase1_batch_llm_enabled: bool = True
|
||
memoir_phase1_batch_llm_max_tokens: int = Field(default=4096, ge=512, le=32_768)
|
||
#: Phase1 批处理 LLM：单次请求最多包含的 segment 数（多块合并，避免 completion 顶满截断）
|
||
memoir_phase1_batch_llm_chunk_size: int = Field(default=24, ge=1, le=500)
|
||
#: 回忆录流水线细粒度进度 Redis 快照 TTL（memoir_pipeline_run:*）
|
||
memoir_pipeline_run_ttl_seconds: int = Field(default=172_800, ge=3600, le=2_592_000)
|
||
# Memoir agents：`invoke_json_object` / `llm_json_call` 的 max_tokens（原硬编码迁至配置）
|
||
memoir_extraction_max_tokens: int = Field(default=1024, ge=64, le=8192)
|
||
memoir_classification_max_tokens: int = Field(default=256, ge=32, le=4096)
|
||
memoir_narrative_max_tokens: int = Field(default=4096, ge=256, le=32_768)
|
||
memoir_narrative_merge_max_tokens: int = Field(default=8192, ge=256, le=64_000)
|
||
memoir_title_max_tokens: int = Field(default=256, ge=32, le=4096)
|
||
memoir_story_route_max_tokens: int = Field(default=1024, ge=64, le=8192)
|
||
memoir_story_batch_plan_max_tokens: int = Field(default=4096, ge=256, le=32_768)
|
||
# 资料抽取（ProfileAgent JSON 模式）
|
||
chat_profile_extract_max_tokens: int = Field(default=512, ge=64, le=4096)
|
||
|
||
# ── ASR ───────────────────────────────────────────────────
|
||
asr_provider: str = "whisper"
|
||
asr_model_size: str = "small"
|
||
asr_device: str = "auto"
|
||
asr_compute_type: str = "auto"
|
||
asr_model_cache_dir: str = ""
|
||
|
||
# ── Tencent SMS ──────────────────────────────────────────
|
||
tencent_sms_secret_id: str = ""
|
||
tencent_sms_secret_key: str = ""
|
||
tencent_sms_sdk_app_id: str = ""
|
||
tencent_sms_sign_name: str = ""
|
||
tencent_sms_template_id: str = ""
|
||
tencent_sms_template_param_count: int = 2
|
||
|
||
# ── Tencent ASR / TTS（共用 Secret；与短信、COS 密钥独立）────────────────
|
||
tencent_secret_id: str = ""
|
||
tencent_secret_key: str = ""
|
||
|
||
# ── TTS (openai | tencent)，与 ASR 独立：仅控制回复侧语音合成 ──
|
||
enable_tts: bool = True
|
||
tts_provider: str = "tencent"
|
||
openai_api_key: str = ""
|
||
tts_voice_type: int = 502001 # Tencent 音色 ID，见 https://cloud.tencent.com/document/product/1073/92668
|
||
tts_codec: str = "mp3"
|
||
|
||
# ── WeChat Pay ───────────────────────────────────────────
|
||
wechat_pay_app_id: str = ""
|
||
wechat_pay_mch_id: str = ""
|
||
wechat_pay_api_v3_key: str = ""
|
||
wechat_pay_private_key_path: str = "certs/apiclient_key.pem"
|
||
wechat_pay_private_key: str = "" # PEM 内容，与 private_key_path 二选一
|
||
wechat_pay_cert_serial_no: str = ""
|
||
wechat_pay_notify_url: str = ""
|
||
wechat_pay_platform_public_key: str = ""
|
||
wechat_pay_platform_public_key_path: str = ""
|
||
wechat_pay_platform_public_key_id: str = ""
|
||
|
||
# ── Alipay ───────────────────────────────────────────────
|
||
alipay_app_id: str = ""
|
||
alipay_private_key: str = ""
|
||
alipay_public_key: str = ""
|
||
alipay_notify_url: str = ""
|
||
alipay_sign_type: str = "RSA2"
|
||
alipay_under_development: str = "true" # "1"/"true"/"yes" 视为开发中不可用
|
||
|
||
# ── Logging ──────────────────────────────────────────────
|
||
# 环境变量 LOG_LEVEL；控制 loguru sink 最低级别（TRACE/DEBUG/INFO/…）
|
||
log_level: str = "INFO"
|
||
# LOG_AGENT_VERBOSE：为 True 时额外输出 Agent 单行 INFO 摘要（耗时、规模），无需全局 DEBUG
|
||
log_agent_verbose: bool = False
|
||
# AGENT_LOG_MAX_CHARS：DEBUG 下记录 prompt/响应预览时的最大字符数；0=不截断（完整输出，慎用）
|
||
agent_log_max_chars: int = Field(default=4096, ge=0, le=50_000_000)
|
||
# AGENT_LOG_OMIT_SYSTEM_MESSAGE_BODY：DEBUG 下访谈/资料聊天日志省略 System 正文（仅 len+sha12）
|
||
agent_log_omit_system_message_body: bool = True
|
||
# AGENT_LOG_JSON_PROMPT_PREFIX_CHARS：DEBUG 下 *.prompt 总长超过下项时再跳过前 N 字符后预览（0=不跳过）
|
||
agent_log_json_prompt_prefix_chars: int = Field(default=0, ge=0, le=500_000)
|
||
# AGENT_LOG_JSON_PROMPT_PREFIX_ONLY_IF_LEN_GT：触发“跳过前缀”的最小 prompt 长度
|
||
agent_log_json_prompt_prefix_only_if_len_gt: int = Field(
|
||
default=4000, ge=0, le=2_000_000
|
||
)
|
||
# AGENT_LOG_PROMPT_MODE：DEBUG 下 *.prompt 记录方式 preview=截断预览 | hash_only=仅 sha12+长度（无正文）
|
||
agent_log_prompt_mode: str = Field(default="preview")
|
||
# AGENT_LOG_PROMPT_DEDUP：DEBUG 下同一 label 连续相同全文时第二条起跳过（减重复模板噪音）
|
||
agent_log_prompt_dedup: bool = False
|
||
# 第三方 stdlib logging（空=自动：DEBUG/TRACE 时 Celery→INFO；否则 Celery 与 httpx 默认 WARNING）
|
||
celery_log_level: str = ""
|
||
httpx_log_level: str = ""
|
||
# 非空时额外写入 JSONL（serialize=True），便于 Loki/ELK；与 stderr 彩色控制台并存
|
||
log_json_file: str = ""
|
||
|
||
@field_validator("celery_purge_broker_on_startup", mode="before")
|
||
@classmethod
|
||
def _coerce_celery_purge_broker_on_startup(cls, v: object) -> bool:
|
||
if isinstance(v, bool):
|
||
return v
|
||
if v is None:
|
||
return False
|
||
return str(v).strip().lower() in ("1", "true", "yes", "on")
|
||
|
||
@field_validator("mock_sms_login_enabled", mode="before")
|
||
@classmethod
|
||
def _coerce_mock_sms_login_enabled(cls, v: object) -> bool:
|
||
if isinstance(v, bool):
|
||
return v
|
||
if v is None:
|
||
return False
|
||
return str(v).strip().lower() in ("1", "true", "yes", "on")
|
||
|
||
@field_validator("log_agent_verbose", mode="before")
|
||
@classmethod
|
||
def _coerce_log_agent_verbose(cls, v: object) -> bool:
|
||
if isinstance(v, bool):
|
||
return v
|
||
if v is None:
|
||
return False
|
||
return str(v).strip().lower() in ("1", "true", "yes", "on")
|
||
|
||
@field_validator("agent_log_omit_system_message_body", mode="before")
|
||
@classmethod
|
||
def _coerce_agent_log_omit_system_message_body(cls, v: object) -> bool:
|
||
if isinstance(v, bool):
|
||
return v
|
||
if v is None:
|
||
return True
|
||
s = str(v).strip().lower()
|
||
if s in ("0", "false", "no", "off"):
|
||
return False
|
||
return True
|
||
|
||
@field_validator("agent_log_prompt_mode", mode="before")
|
||
@classmethod
|
||
def _normalize_agent_log_prompt_mode(cls, v: object) -> str:
|
||
if v is None:
|
||
return "preview"
|
||
s = str(v).strip().lower()
|
||
if s not in ("preview", "hash_only"):
|
||
return "preview"
|
||
return s
|
||
|
||
@field_validator("agent_log_prompt_dedup", mode="before")
|
||
@classmethod
|
||
def _coerce_agent_log_prompt_dedup(cls, v: object) -> bool:
|
||
if isinstance(v, bool):
|
||
return v
|
||
if v is None:
|
||
return False
|
||
return str(v).strip().lower() in ("1", "true", "yes", "on")
|
||
|
||
# ── Misc ─────────────────────────────────────────────────
|
||
enable_test_subscription: int = 0
|
||
enable_test_plan: str = "" # "1" / "true" / "yes" 为 True
|
||
enable_docs: bool = True
|
||
|
||
# ── Memoir Image ─────────────────────────────────────────
|
||
memoir_image_enabled: bool = False
|
||
# True：图片 LLM prompt 失败时不使用英语降级模板（需产品与任务失败流确认后开启）
|
||
image_prompt_fallback_disabled: bool = False
|
||
memoir_image_poll_interval: int = 3
|
||
memoir_image_max_attempts: int = 20
|
||
memoir_image_provider: str = "liblib"
|
||
memoir_image_style_default: str = "watercolor"
|
||
memoir_image_size_default: str = "1280x720"
|
||
memoir_image_download_hosts: str = ""
|
||
# 章节 canonical_markdown 中至少含多少张 asset:// 正文插图才生成/展示章节封面（≥ 该值即满足；0 表示不以此条件拦截）
|
||
memoir_min_inline_images_for_chapter_cover: int = Field(default=1, ge=0, le=100)
|
||
# Story 正文至少多少字才创建主图 intent / 调图（0 表示不限制）
|
||
story_image_min_body_chars: int = 400
|
||
# generate_story_image 入队去重（Redis SET NX，秒）
|
||
story_image_enqueue_dedup_ttl: int = Field(default=300, ge=30, le=86400)
|
||
# 章节物化异步任务延迟入队（秒），削峰
|
||
recompose_chapter_delay_seconds: int = Field(default=8, ge=0, le=600)
|
||
# 与 memoir pipeline 一致的章节互斥锁 TTL（秒）；应覆盖 Phase2 / recompose 的 P95 时长
|
||
chapter_pipeline_lock_ttl_seconds: int = Field(default=360, ge=10, le=3600)
|
||
# Append 硬上限：canonical 字符数、版本数（超限强制 new_story）
|
||
story_append_max_canonical_chars: int = Field(default=12000, ge=1000, le=500_000)
|
||
story_append_max_versions: int = Field(default=20, ge=1, le=500)
|
||
# StoryRouteAgent：候选 JSON 预算（保守默认，可调大）
|
||
story_route_candidate_body_max_chars: int = Field(default=2200, ge=200, le=8000)
|
||
story_route_candidate_total_max_chars: int = Field(
|
||
default=20_000, ge=2000, le=100_000
|
||
)
|
||
story_route_long_body_head_chars: int = Field(default=700, ge=100, le=4000)
|
||
story_route_long_body_tail_chars: int = Field(default=700, ge=100, le=4000)
|
||
story_route_summary_min_chars: int = Field(default=30, ge=0, le=500)
|
||
story_route_index_preview_chars: int = Field(default=140, ge=20, le=500)
|
||
# 童年/求学/家庭：本批口述低于该字数且路由为 new 时，倾向续写到默认候选，减少碎篇
|
||
memoir_story_route_append_guardrail_oral_chars: int = Field(
|
||
default=1800, ge=0, le=50_000
|
||
)
|
||
# Evidence 检索 top_k：大批次 unit 时降低检索量
|
||
evidence_top_k_default: int = Field(default=10, ge=1, le=50)
|
||
evidence_top_k_large_batch: int = Field(default=5, ge=1, le=50)
|
||
evidence_large_batch_threshold: int = Field(default=3, ge=1, le=100)
|
||
# Story/Chapter 标题在正文达到此字数后才由 LLM 生成；之前用占位符
|
||
story_title_min_body_chars: int = Field(default=60, ge=0, le=10_000)
|
||
# 回忆录 Celery：累计 strip 后口述字数未达此值则暂缓提交（0=关闭，仅防抖后提交）
|
||
memoir_segment_batch_min_chars: int = Field(default=50, ge=0, le=50_000)
|
||
# 本批首条 segment 入队起最长等待（秒），超时则提交（即使字数不足）
|
||
memoir_segment_batch_max_wait_seconds: float = Field(
|
||
default=60.0, ge=0.0, le=3600.0
|
||
)
|
||
# 回忆录叙事 Phase 2（ Celery）触发：单条口述达到该 strip 字数则立即跑叙事
|
||
memoir_narrative_immediate_char_threshold: int = Field(default=50, ge=0, le=50_000)
|
||
# 同一 topic_category 下未叙事段数达到该值则触发 Phase 2
|
||
memoir_narrative_batch_min_segments: int = Field(default=3, ge=1, le=500)
|
||
# 同上，累计 user_input_text 字符数（strip 后由 Segment 列 length 近似）
|
||
memoir_narrative_batch_min_chars: int = Field(default=80, ge=0, le=500_000)
|
||
# Phase 1 完成后未触发 Phase 2 时，延迟任务兜底（秒）；新 Phase 1 会 revoke 旧定时
|
||
memoir_narrative_batch_max_wait_seconds: float = Field(
|
||
default=120.0, ge=1.0, le=3600.0
|
||
)
|
||
# False：Celery/批处理更新 slot 时不改写 MemoirState.current_stage（访谈路径仍可由 switch_stage 推进）
|
||
# True：仅当 chat_bucket( proposed ) == chat_bucket( existing ) 时允许批处理对齐 current_stage
|
||
memoir_extraction_updates_current_stage: bool = False
|
||
# True：FidelityCheckAgent JSON/LLM 解析失败时放行（仅建议 append 场景配合 existing 兜底）
|
||
memoir_fidelity_fail_open_on_parse_error: bool = False
|
||
# 正文与 evidence 文本的最长公共子串达到该长度且 oral/旧正文未覆盖时，回退为安全正文
|
||
memoir_narrative_evidence_overlap_min_chars: int = Field(default=14, ge=8, le=256)
|
||
# True：启用短「场合锚点」词检测（聚餐/那晚等），须同时在摘录中出现且口述未覆盖才回退
|
||
memoir_evidence_scene_anchor_check_enabled: bool = True
|
||
# True：标题生成时 slots 仅保留在 oral 或正文摘录中出现的条目（减少档案串台）
|
||
memoir_title_slots_require_body_or_oral_match: bool = True
|
||
# True：标题中出现高置信「履历链」短语则须在 hay（正文+口述+已传 slots）中有逐字依据，否则降级占位
|
||
memoir_title_hay_grounding_strict_phrases_enabled: bool = True
|
||
# True：章节物化拿不到 pipeline 锁时 Celery retry（避免长期跳过导致 dirty 不收敛）
|
||
memoir_recompose_retry_on_lock_contention: bool = True
|
||
# Phase2 立即派发使用固定 task_id，减少同类目重复入队（超时任务仍用独立 id）
|
||
memoir_phase2_singleflight_immediate: bool = True
|
||
# True：Phase2 路由低置信（no_llm/parse_error/invalid_target）时不写 Story，
|
||
# 把 segment 标记为 narrative_deferred_until 之后再重试。
|
||
memoir_route_defer_enabled: bool = True
|
||
# 低置信延迟时长（秒）：到期前不消费这些 segment，避免后台空转
|
||
memoir_route_defer_seconds: float = Field(default=120.0, ge=1.0, le=3600.0)
|
||
# 同一类目最多自动延迟次数；达到上限后 segment 仅靠新素材到达激活，不再自动重试
|
||
memoir_route_defer_max_attempts: int = Field(default=3, ge=1, le=20)
|
||
# True：Phase2 首稿后异步运行质量增强（fidelity recheck、标题润色、LLM 归一）
|
||
memoir_quality_pass_enabled: bool = True
|
||
memoir_quality_pass_delay_seconds: int = Field(default=5, ge=0, le=300)
|
||
|
||
# ── Memory 检索与富化 ─────────────────────────────────────
|
||
# False：跳过 ingest 后 LLM 富化（摘要/事实/时间线）
|
||
memory_enrichment_enabled: bool = True
|
||
memory_enrichment_max_chars: int = Field(default=12000, ge=1000, le=100_000)
|
||
|
||
# ── Memory compaction（近重复 chunk 软排除；事件触发 + Redis 防抖 + 用户锁；需 worker + Beat 跑 sweep）──
|
||
memory_compaction_enabled: bool = True
|
||
memory_compaction_debounce_seconds: int = Field(default=105, ge=10, le=3600)
|
||
memory_compaction_lock_ttl_seconds: int = Field(default=600, ge=60, le=7200)
|
||
memory_compaction_chunk_similarity_threshold: float = Field(
|
||
default=0.92, ge=0.5, le=0.999
|
||
)
|
||
memory_compaction_min_layers_for_exclude: int = Field(default=2, ge=1, le=3)
|
||
memory_compaction_max_chunks_per_run: int = Field(default=200, ge=1, le=10_000)
|
||
memory_compaction_max_excludes_per_run: int = Field(default=50, ge=1, le=1000)
|
||
memory_compaction_max_neighbors_per_chunk: int = Field(default=25, ge=5, le=100)
|
||
memory_compaction_text_jaccard_min: float = Field(default=0.55, ge=0.0, le=1.0)
|
||
memory_compaction_metadata_event_year_window: int = Field(default=1, ge=0, le=50)
|
||
# Beat sweep：扫描最近 N 小时内有新 chunk 的用户并调度 compaction
|
||
memory_compaction_sweep_recent_hours: int = Field(default=24, ge=1, le=168)
|
||
|
||
# ── Liblib ───────────────────────────────────────────────
|
||
liblib_access_key: str = ""
|
||
liblib_secret_key: str = ""
|
||
liblib_base_url: str = "https://openapi.liblibai.cloud"
|
||
liblib_template_uuid: str = ""
|
||
|
||
# ── Tencent COS ──────────────────────────────────────────
|
||
tencent_cos_secret_id: str = ""
|
||
tencent_cos_secret_key: str = ""
|
||
tencent_cos_region: str = "ap-shanghai"
|
||
tencent_cos_bucket: str = ""
|
||
tencent_cos_base_url: str = ""
|
||
tencent_cos_token: str = ""
|
||
|
||
# ── Internal regression evaluation lab（独立入口，不挂在消费者 API）────
|
||
internal_eval_api_key: str = ""
|
||
internal_eval_enable_docs: bool = False
|
||
# 逗号分隔；空则内部 API 不额外限制 Origin（仍可依赖 internal_eval_api_key）
|
||
internal_eval_cors_origins: str = ""
|
||
# 智谱 GLM-5：评审模型（OpenAI 兼容 Chat Completions，与 langchain-openai 一致）
|
||
eval_judge_api_key: str = ""
|
||
eval_judge_base_url: str = "https://open.bigmodel.cn/api/paas/v4"
|
||
eval_judge_model: str = "glm-5"
|
||
eval_judge_temperature: float = 0.3
|
||
# 评测评审：DeepSeek（OpenAI 兼容）；默认 deepseek-v4-flash + 非思考（对齐定价页非思考用法；非 v4-pro）
|
||
eval_judge_deepseek_model: str = "deepseek-v4-flash"
|
||
# 当仅指定 deepseek-v4-flash、未用弃用名区分时，是否走思考模式（与 eval_judge_deepseek_model 联用）
|
||
eval_judge_deepseek_thinking_enabled: bool = False
|
||
eval_judge_deepseek_context_window_tokens: int = Field(
|
||
default=64_000,
|
||
ge=4096,
|
||
le=2_000_000,
|
||
description="DeepSeek 评审专用上下文预算（用于 transcript 截断；与 GLM 200K 分离）",
|
||
)
|
||
# GLM-5 输入上下文 200K（https://docs.bigmodel.cn）
|
||
eval_judge_context_window_tokens: int = Field(
|
||
default=200_000, ge=4096, le=2_000_000
|
||
)
|
||
# 预留给完成 tokens（json 输出）及路由误差
|
||
eval_judge_completion_reserve_tokens: int = Field(default=4096, ge=256, le=131_072)
|
||
eval_judge_prompt_budget_safety_tokens: int = Field(default=2048, ge=0, le=32_768)
|
||
# transcript 混合中英文时 token/字 估值（略低于 1.2 可多给汉字篇幅；若评审请求被拒可回调高）
|
||
eval_judge_approx_tokens_per_char: float = Field(default=1.0, ge=0.3, le=8.0)
|
||
# 整段/逐轮节选 transcript 最大字符；0=按 eval_judge_context_window_tokens 自动扣 rubric 头
|
||
eval_judge_max_transcript_chars: int = Field(default=0, ge=0, le=2_000_000)
|
||
# 双 transcript 对比流：每条对话上限；0=按上下文平分（扣 overhead）
|
||
eval_judge_max_compare_transcript_chars_each: int = Field(
|
||
default=0, ge=0, le=2_000_000
|
||
)
|
||
# 对比 prompt 固定开销（模板 + 两份评分 JSON）的字符估值；略低则 transcript 合计空间更大
|
||
eval_judge_compare_prompt_overhead_chars: int = Field(
|
||
default=10_000, ge=500, le=500_000
|
||
)
|
||
# 回忆录音评：章节 LLM 并发上限（仅评审请求；准备阶段仍串行访问 DB）
|
||
eval_judge_memoir_chapter_concurrency: int = Field(
|
||
default=4,
|
||
ge=1,
|
||
le=32,
|
||
)
|
||
# 回忆录评审 prompt 内粗截断（汉字计字符）；万字级章节请保持 body ≥ 正文峰值
|
||
eval_judge_memoir_body_max_chars: int = Field(
|
||
default=36_000,
|
||
ge=8_000,
|
||
le=500_000,
|
||
description="【当前回忆录正文】注入评审 prompt 前的最大字符",
|
||
)
|
||
eval_judge_memoir_evidence_max_chars: int = Field(
|
||
default=32_000,
|
||
ge=8_000,
|
||
le=500_000,
|
||
description="对话证据 / 结构化证据 / 参考基线各块的最大字符（与 eval_trace_format 对齐）",
|
||
)
|
||
# json_object 完成预算；MemoirJudgeOutput 字段多，需预留足量 token
|
||
eval_judge_memoir_completion_max_tokens: int = Field(
|
||
default=3072,
|
||
ge=512,
|
||
le=16_384,
|
||
)
|
||
# 候选对话回放：与生产访谈类似的温度
|
||
eval_candidate_temperature: float = 0.7
|
||
# 门禁：受保护 session 合成份数下跌超过该阈值视为回归（0–100 分制）
|
||
eval_gate_protected_regression_threshold: float = Field(
|
||
default=2.0, ge=0.0, le=100.0
|
||
)
|
||
# 执行 LLM 判分与回放（Celery 未跑时可关，仅跑结构/导入）
|
||
eval_execution_enabled: bool = True
|
||
|
||
|
||
settings = Settings()
|