Spaces:
Running
Running
File size: 14,103 Bytes
50efdc6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 | """
src/preprocess/email_preprocessor.py
LangGraph ์ง์
์ ์ด๋ฉ์ผ ์ ์ฒ๋ฆฌ โ ์์ฐ์ด ์ด๋ฉ์ผ โ ๊ตฌ์กฐํ๋ EmailContext.
๋ชฉ์ :
ํ์ฌ router / worker_a / worker_b / synthesizer 4๊ฐ ๋
ธ๋๊ฐ ๊ฐ์ LLM ํธ์ถ๋ก
๊ฐ์ ์ด๋ฉ์ผ์ ๋ค์ ํ์ฑํ๋ค(๋๊ฐ ๋ณด๋๋์ง, ๋ฌด์์ ์์ฒญํ๋์ง, ์ด๋ค ์ซ์๊ฐ ์๋์งโฆ).
์ด ์ ์ฒ๋ฆฌ๋ฅผ ํ ๋ฒ ์ํํด์ EmailContext์ ๋ด์๋๋ฉด:
1) ๋ค์ด์คํธ๋ฆผ ๋
ธ๋์ LLM ํธ์ถ์ด ์งง์ hint๋ฅผ ๋ฐ์ ์ ํ๋/์ผ๊ด์ฑ ํฅ์
2) ๊ฐ์ ์ถ๋ก ์ 4๋ฒ ๋ฐ๋ณตํ๋ ๋ญ๋น ๊ฐ์
3) ๋๋ฒ๊น
๊ฐ์์ฑ ํฅ์ (์ด๋ฉ์ผ ํ ํต์ด ์ด๋ป๊ฒ ํ์ฑ๋๋์ง ํ๊ณณ์์ ํ์ธ)
์ค๊ณ:
- LangGraph ์ธ๋ถ์์ ๋์ (graph_builder / api/server / eval ์ง์
์ ์์ ํธ์ถ).
- ๊ฒฐ๊ณผ๋ AgentState["email_context"]์ dict๋ก ์ ์ฌ. ๋
ธ๋๋ค์ ์์ผ๋ฉด ํ์ฉ, ์์ผ๋ฉด
๊ธฐ์กด ๋์(์๋ณธ user_input ํ์ฑ)์ผ๋ก ์๋ fallback โ ํ์ํธํ.
- ์คํจํด๋ ๊ทธ๋ํ ํ๋ฆ์ ์ค๋จํ์ง ๋ง ๊ฒ. EmailContext.preprocess_ok=False๋ก ํ์๋ง.
"""
from __future__ import annotations
import logging
import os
import re
from typing import Optional
from dotenv import load_dotenv
load_dotenv()
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, Field
from src.config import get_config
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# ๊ตฌ์กฐํ๋ ๊ฒฐ๊ณผ ์คํค๋ง
# ---------------------------------------------------------------------------
class EmailContext(BaseModel):
"""์ ์ฒ๋ฆฌ ๊ฒฐ๊ณผ โ LangGraph ๋
ธ๋๋ค์ด user_input๊ณผ ํจ๊ป ์ฐธ๊ณ ํ๋ ๊ตฌ์กฐํ ์ ๋ณด."""
# ์๋ฌธ (๊ทธ๋๋ก ๋ณด์กด โ ๋ค์ด์คํธ๋ฆผ ๋
ธ๋๊ฐ ํ์ ์ ์ฐธ๊ณ )
raw_text: str = Field(..., description="์๋ณธ ์ด๋ฉ์ผ ํ
์คํธ (์์ ์์)")
# ๋ฉํ๋ฐ์ดํฐ (LLM์ด ์ถ์ถ, ์์ผ๋ฉด ๋น ๋ฌธ์์ด)
sender_name: str = Field("", description="๋ฐ์ ์ ์ด๋ฆ (์๋ช
/From ๋ผ์ธ ๋ฑ์์ ์ถ์ถ). ๋ชจ๋ฅด๋ฉด ๋น ๋ฌธ์์ด.")
sender_email: str = Field("", description="๋ฐ์ ์ ์ด๋ฉ์ผ ์ฃผ์. ๋ชจ๋ฅด๋ฉด ๋น ๋ฌธ์์ด.")
sender_company: str = Field("", description="๋ฐ์ ์ ์์ ํ์ฌ๋ช
. ๋ชจ๋ฅด๋ฉด ๋น ๋ฌธ์์ด.")
recipient: str = Field("", description="์์ ์ (๋ด๋น์/ํ๋ช
). ๋ชจ๋ฅด๋ฉด ๋น ๋ฌธ์์ด.")
subject: str = Field("", description="์ ๋ชฉ. ๋ณธ๋ฌธ์์ ์ถ์ ํด๋ ๋จ.")
language: str = Field("en", description="์ด๋ฉ์ผ ์ธ์ด (en/ko/ja/zh/...). ๊ธฐ๋ณธ en.")
# ์ ์ ๋ ๋ณธ๋ฌธ โ ์ธ์ฌ๋ง/์๋ช
/Disclaimer ์ ์ธ, ๋ณธ๋ฌธ ํต์ฌ๋ง
cleaned_body: str = Field(..., description="์ธ์ฌยท์๋ช
ยท๋ฉด์ฑ
์กฐํญ์ ์ ์ธํ ๋ณธ๋ฌธ ํต์ฌ")
# ์์ฒญ ์์ฝ โ ํ๋ ๋ฌธ์ฅ
request_summary: str = Field(
...,
description="์ด๋ฉ์ผ์ด ๋ฌด์์ ์์ฒญํ๋์ง ํ๋ ๋ฌธ์ฅ ์์ฝ (์์ด๋ก). "
"์: 'Change quantity of item 10 on order 4500023456 to 100 units.'",
)
# ์ง์ ์ง๋ฌธ๋ง ๋ถ๋ฆฌยท์ ๊ทํ โ RAG(worker_b) ๊ฒ์ ์ฟผ๋ฆฌ ์
๋ ฅ์ผ๋ก ์ฌ์ฉ
question_summary: str = Field(
"",
description="์ด๋ฉ์ผ์ SAP ์ง์/๋ฐฉ๋ฒ(how-to/์ ์ฑ
/๊ฐ๋
) ์ง๋ฌธ์ด ์์ผ๋ฉด, ERP ์ก์
๋งฅ๋ฝ์ "
"๋ชจ๋ ์ ๊ฑฐํ๊ณ ํ๋์ ๋ช
ํํ๊ณ ์๊ธฐ์๊ฒฐ์ ์ธ ์์ด ์ง๋ฌธ์ผ๋ก ์ฌ์์ฑ. ํน์ ์ฃผ๋ฌธ/์์ดํ
"
"๋ฒํธ๋ ์๋ ๊ฐ์ ํธ๋์ญ์
๋ํ
์ผ์ ๋นผ๊ณ ์ผ๋ฐํํ๋ค. ์ง์ ์ง๋ฌธ์ด ์์ผ๋ฉด ๋น ๋ฌธ์์ด. "
"์: 'How do I view the incompletion log for a sales order before creating an outbound delivery?'",
)
# โโ ๋น ๋ฅธ ๋ผ์ฐํ
ํํธ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
mentions_action: bool = Field(
False,
description="ERP ์์ ์์ฒญ(์๋/๋ ์ง/์ทจ์/์ฃผ์ ๋ณ๊ฒฝ ๋ฑ)์ด ๋ช
์๋์ด ์๋๊ฐ",
)
mentions_question: bool = Field(
False,
description="SAP/์ ์ฑ
/๋งค๋ด์ผ์ ๋ํ ์ง์/๋ฌธ์ ์ง๋ฌธ์ด ๋ช
์๋์ด ์๋๊ฐ",
)
# โโ ์ฌ์ ์ถ์ถ๋ ํต์ฌ ์ํฐํฐ (๋ค์ด์คํธ๋ฆผ์ด hint๋ก ํ์ฉ) โโโโโโโโโโโโโโโโโโโโ
order_ids: list[str] = Field(
default_factory=list,
description="์ด๋ฉ์ผ์ ๋ช
์๋ ๋ชจ๋ ์์
์ค๋ ๋ฒํธ(VBELN) ํ๋ณด โ ์๋ณธ ์ซ์ ๊ทธ๋๋ก, ํจ๋ฉ ์์ด",
)
item_nos: list[str] = Field(
default_factory=list,
description="์ด๋ฉ์ผ์ ๋ช
์๋ ๋ชจ๋ ์์ดํ
๋ฒํธ(POSNR) ํ๋ณด โ ์๋ณธ ์ซ์ ๊ทธ๋๋ก",
)
# โโ ๋ฉํ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
preprocess_ok: bool = Field(True, description="์ ์ฒ๋ฆฌ ์ฑ๊ณต ์ฌ๋ถ (LLM ์คํจ ์ False)")
error: str = Field("", description="์คํจ ์ฌ์ (preprocess_ok=False์ผ ๋๋ง)")
# ---------------------------------------------------------------------------
# Prompt
# ---------------------------------------------------------------------------
_SYSTEM_PROMPT = """\
You are a B2B email preprocessor for an SAP ERP support pipeline. Your job is to
read a single customer email and produce a structured summary that downstream
agents (router, ERP action extractor, RAG question extractor, reply composer)
will consume.
Extract the following fields from the email and output ONLY valid JSON matching
the required schema.
Field guidance:
โข sender_name / sender_email / sender_company:
Look in the signature block ("Best regards, John Smith / john@acme.com /
Acme Corp"), or in "From:" headers if present. If a field is not
recoverable from the email, return an empty string. Do NOT invent.
โข recipient:
The addressed party ("Hi Support Team", "Dear SAP Desk"). Empty if unclear.
โข subject:
The Subject: line if present, otherwise a short topic phrase you infer
from the body (under 80 chars).
โข language:
Two-letter code: "en", "ko", "ja", "zh", "de", "es", "fr". Default "en".
โข cleaned_body:
The substantive body โ strip greetings ("Dear ..."), sign-offs ("Best
regards, ..."), legal disclaimers, and "Sent from my iPhone"-style
footers. Keep the actual request and any context the customer provides.
Preserve order numbers, item numbers, quantities, dates exactly as
written.
โข request_summary:
ONE or TWO sentences in ENGLISH that describe what the customer wants.
Be specific โ include order number, item number, action, and quantities
or dates if mentioned. Examples:
"Change quantity of item 10 on order 4500023456 to 100 units."
"Reduce quantity of item 20 on order 6105 by 50, and ask about the
late-delivery penalty policy."
โข question_summary:
If (and only if) the email contains an SAP knowledge / how-to / policy /
concept question, restate THAT question as ONE clear, self-contained
question in ENGLISH โ optimized as a documentation search query.
Remove ONLY the transaction instance data (order/item numbers, quantities,
dates) and the action request itself. Otherwise stay close to the
customer's own wording: PRESERVE, verbatim, every SAP term, feature /
screen / field name, and concept keyword they used โ these are the search
keys the retriever matches on. Do NOT paraphrase a specific SAP term into a
generic one, and do NOT swap in a related but different concept for the one
asked about. If the question has two distinct parts, keep both. If there is
no knowledge question, return an empty string.
โข mentions_action:
true if the email asks to MODIFY anything in the ERP โ change quantity,
change delivery date, cancel an item, change shipping address, unblock a
delivery, change carrier/batch/payment terms. false if it is a pure
knowledge question.
โข mentions_question:
true if the email asks about HOW SAP works / what a policy says / how to
configure something / definitions โ i.e. knowledge-base questions answered
from documentation. false if it is purely a transaction request.
mentions_action and mentions_question are NOT mutually exclusive โ an
email can have both.
โข order_ids:
EVERY distinct order number (VBELN) mentioned. Copy digits EXACTLY as
written โ do NOT zero-pad, do NOT add or drop digits. Numbers may
appear after "order", "order #", "sales order", "PO". Return as strings,
no commas/dashes. Empty list if none.
โข item_nos:
EVERY distinct item / line / position number (POSNR) mentioned. Copy
digits EXACTLY. Numbers may appear after "item", "line", "position".
Empty list if none.
CRITICAL โ number handling:
A single email often contains the order number, the item number, AND a
quantity. They are DIFFERENT things. Map each to the correct field. Never
put a quantity ("100 units", "50 pcs") into order_ids or item_nos.
Output ONLY valid JSON matching the schema. Set raw_text to the EXACT input
email (unchanged). Set preprocess_ok=true and error="" โ these are only set to
false by the calling code on failure.
"""
_HUMAN_TEMPLATE = "Email to preprocess:\n\n{email}"
_PROMPT = ChatPromptTemplate.from_messages([
("system", _SYSTEM_PROMPT),
("human", _HUMAN_TEMPLATE),
])
# ---------------------------------------------------------------------------
# LLM factory (OpenRouter)
# ---------------------------------------------------------------------------
def _build_llm() -> ChatOpenAI:
"""์ ์ฒ๋ฆฌ LLM. configs.yaml์ models.preprocessor๋ฅผ ์ฐ์ ์ฌ์ฉํ๊ณ ,
์ค์ ์ด ์์ผ๋ฉด worker_a(์ด๋ฏธ ๊ตฌ์กฐํ ์ถ์ถ์ ์ ๋์)์ ๋์ผํ ๋ชจ๋ธ๋ก ํด๋ฐฑ."""
cfg = get_config()
# ๋์ lookup โ ์ ์ค์ ํค๊ฐ ์์ด๋ ๋์ํ๋๋ก
preproc_cfg = getattr(cfg.models, "preprocessor", None) or cfg.models.worker_a
api_key = os.getenv("OPENROUTER_API_KEY", "")
if not api_key:
raise EnvironmentError("OPENROUTER_API_KEY is not set. Check your .env file.")
return ChatOpenAI(
model=preproc_cfg.name,
temperature=preproc_cfg.temperature,
openai_api_key=api_key,
openai_api_base=cfg.openrouter.base_url,
default_headers={
"HTTP-Referer": "https://github.com/daisysooyeon/SAP-ERP-AI-Agent",
"X-Title": "SAP ERP AI Agent - Preprocessor",
},
)
_llm: Optional[ChatOpenAI] = None
_chain = None
def _get_chain():
"""LLM ์ฒด์ธ์ lazyํ๊ฒ ์์ฑ โ import ์์ ์ API ํค ๊ฒ์ฆ์ ๋ฏธ๋ฃจ๊ธฐ ์ํจ."""
global _llm, _chain
if _chain is None:
_llm = _build_llm()
_chain = _PROMPT | _llm.with_structured_output(EmailContext)
return _chain
# ---------------------------------------------------------------------------
# Regex ๊ธฐ๋ฐ backstop entity ์ถ์ถ (LLM ์คํจ/๋๋ฝ ๋ณด์กฐ)
# ---------------------------------------------------------------------------
_ORDER_RE = re.compile(
r"(?:order|order\s*#|sales\s+order|PO)\s*[:#]?\s*(\d{3,10})",
re.IGNORECASE,
)
_ITEM_RE = re.compile(
r"(?:item|line|position)\s*[:#]?\s*(\d{1,6})",
re.IGNORECASE,
)
def _regex_entities(text: str) -> tuple[list[str], list[str]]:
"""LLM์ด ์คํจํ๊ฑฐ๋ ๋น ๋จ๋ฆด ๋๋ฅผ ๋๋นํ ์ ๊ท์ backstop. ์ค๋ณต ์ ๊ฑฐ."""
order_ids = list(dict.fromkeys(_ORDER_RE.findall(text)))
item_nos = list(dict.fromkeys(_ITEM_RE.findall(text)))
return order_ids, item_nos
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def preprocess_email(raw_email: str) -> EmailContext:
"""์ด๋ฉ์ผ์ LLM ํ ๋ฒ ํธ์ถ๋ก ๊ตฌ์กฐํ๋ EmailContext๋ก ๋ณํ.
LLM์ด ์คํจํด๋ EmailContext.preprocess_ok=False๋ก ํ์ํ๊ณ ์ ๊ท์ backstop์ผ๋ก
์ต์ํ์ ์ ๋ณด๋ฅผ ์ฑ์ ๋ฐํํ๋ค. ํธ์ถ์(graph entry point)๋ ๊ทธ๋ฅ ํต๊ณผ์ํค๋ฉด
๋๊ณ , ๋ค์ด์คํธ๋ฆผ ๋
ธ๋๋ ๋น ํ๋๋ฅผ ๋ณด๋ฉด ๊ธฐ์กด user_input ํ์ฑ์ผ๋ก ํด๋ฐฑํ๋ค.
"""
raw_email = raw_email or ""
logger.info("[preprocess] Preprocessing email (%d chars)โฆ", len(raw_email))
try:
chain = _get_chain()
result: EmailContext = chain.invoke({"email": raw_email})
# LLM์ด raw_text๋ฅผ ๋น์ฐ๊ฑฐ๋ ๋ณํํ์ ์ ์์ผ๋ฏ๋ก ๊ฒฐ์ ๋ก ์ ์ผ๋ก ๋ฎ์ด์ด๋ค
result.raw_text = raw_email
# LLM์ด ์ํฐํฐ๋ฅผ ๋น ๋จ๋ ธ์ผ๋ฉด ์ ๊ท์์ผ๋ก ๋ณด๊ฐ (๋ฎ์ด์ฐ์ง ์๊ณ union)
rx_orders, rx_items = _regex_entities(raw_email)
if not result.order_ids and rx_orders:
result.order_ids = rx_orders
if not result.item_nos and rx_items:
result.item_nos = rx_items
logger.info(
"[preprocess] OK | sender=%r action=%s question=%s "
"orders=%s items=%s lang=%s",
result.sender_name, result.mentions_action,
result.mentions_question, result.order_ids, result.item_nos,
result.language,
)
return result
except Exception as e:
# ์คํจํด๋ ํ๋ฆ์ ๋์ง ์๋๋ค โ backstop์ผ๋ก ์ต์ ์ ๋ณด ์ฑ์ ๋ฐํ
logger.error("[preprocess] FAILED: %s โ falling back to regex backstop only", e)
rx_orders, rx_items = _regex_entities(raw_email)
return EmailContext(
raw_text=raw_email,
cleaned_body=raw_email, # ์๋ณธ์ ๊ทธ๋๋ก ์ ์ง
request_summary="",
order_ids=rx_orders,
item_nos=rx_items,
preprocess_ok=False,
error=f"{type(e).__name__}: {e}",
)
|