File size: 3,374 Bytes
9381502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json
import time
from typing import List, Optional

from pydantic import BaseModel, Field
from langchain.prompts import ChatPromptTemplate
from langchain_core.output_parsers import PydanticOutputParser


class QuestionParsedResult(BaseModel):
    stem: str = Field(description="๋ฌธ์ œ ๋ณธ๋ฌธ", default="")
    options: List[str] = Field(description="๋ณด๊ธฐ ๋ชฉ๋ก (๊ฐ๊ด€์‹์ผ ๊ฒฝ์šฐ)", default_factory=list)
    answer: List[str] = Field(description="์ •๋‹ต ๋˜๋Š” ์„ ํƒ๊ฒฐ๊ณผ", default_factory=list)
    explanation: str = Field(description="๊ฐ„๋‹จํ•œ ํ•ด์„ค", default="")
    question_type: str = Field(description="๋ฌธ์ œ ์œ ํ˜• (mcq, yes_no, sequence, code, scenario)")
    code: Optional[str] = Field(description="์ฝ”๋“œ๊ฐ€ ์žˆ๋‹ค๋ฉด ํฌํ•จ", default="")
    sequence: Optional[List[str]] = Field(description="์ˆœ์„œํ˜•์ผ ๊ฒฝ์šฐ ํ•ญ๋ชฉ", default_factory=list)

parser = PydanticOutputParser(pydantic_object=QuestionParsedResult)

QUESTION_PARSE_PROMPT = ChatPromptTemplate.from_template(
    """
๋„ˆ๋Š” OCR๋กœ ์ธ์‹๋œ ์‹œํ—˜ ๋ฌธ์ œ๋ฅผ ๋ถ„์„ํ•˜์—ฌ ๊ตฌ์กฐํ™”ํ•˜๋Š” ์ „๋ฌธ ํŒŒ์„œ์ด๋‹ค.
๋ฌธ์ œ ์œ ํ˜•์€ ๋‹ค์Œ ์ค‘ ํ•˜๋‚˜์ผ ์ˆ˜ ์žˆ๋‹ค:
- mcq: ๊ฐ๊ด€์‹
- yes_no: ์ฐธ/๊ฑฐ์ง“ ๋˜๋Š” ์˜ˆ/์•„๋‹ˆ์˜ค
- sequence: ์ˆœ์„œํ˜• ๋ฌธ์ œ
- code: ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ๋ฌธ์ œ
- scenario: ์‹œ๋‚˜๋ฆฌ์˜ค ๊ธฐ๋ฐ˜ ๋ฌธ์ œ

์ด ๋ฌธ์ œ๋Š” ์ฃผ๋กœ '{question_type}' ์œ ํ˜•์œผ๋กœ ์ถ”์ •๋œ๋‹ค.

{format_instructions}

๋ฌธ์ œ ์›๋ฌธ:
{ocr_text}
"""
)


def detect_question_type(text: str) -> str:
    lowered = text.lower()
    if any(k in lowered for k in ["์˜ˆ", "์•„๋‹ˆ์˜ค", "yes", "no", "true", "false"]):
        return "yes_no"
    if any(k in lowered for k in ["์ˆœ์„œ", "์ •๋ ฌ", "drag", "drop", "์ˆœ์„œ๋Œ€๋กœ"]):
        return "sequence"
    if any(k in lowered for k in ["json", "{", "}", "az ", "set-az", "cli", "powershell", "bash", "cmd", "์ฝ”๋“œ"]):
        return "code"
    if any(k in lowered for k in ["์‹œ๋‚˜๋ฆฌ์˜ค", "case", "contoso", "litware", "fabrikam", "์กฐ๊ฑด", "์ƒํ™ฉ"]):
        return "scenario"
    return "mcq"


def build_question_parser_chain(llm):
    # LCEL: Prompt -> LLM -> Pydantic Parser
    return QUESTION_PARSE_PROMPT | llm | parser


def parse_question_with_chain(llm, page: int, text: str) -> dict:
    question_type = detect_question_type(text)
    chain = build_question_parser_chain(llm)

    try:
        started_at = time.time()
        # parser.get_format_instructions()๋ฅผ ํ†ตํ•ด ํ”„๋กฌํ”„ํŠธ์— JSON ์Šคํ‚ค๋งˆ๋ฅผ ์ฃผ์ž…ํ•ฉ๋‹ˆ๋‹ค.
        parsed_result = chain.invoke(
            {
                "question_type": question_type,
                "ocr_text": text[:1800],
                "format_instructions": parser.get_format_instructions(),
            }
        )
        
        # Pydantic ๋ชจ๋ธ์„ dict๋กœ ๋ณ€ํ™˜ (Pydantic v1/v2 ํ˜ธํ™˜์„ฑ์„ ์œ„ํ•ด dict() ์‚ฌ์šฉ)
        parsed = parsed_result.dict()
        parsed["page"] = page
        
        print(f"[LLM] p{page} ({question_type}) parsed in {time.time() - started_at:.1f}s")
        return parsed
        
    except Exception as exc:
        print(f"[WARN] LLM parse failed (p{page}): {exc}")
        return {
            "page": page,
            "stem": text[:400],
            "options": [],
            "answer": [],
            "explanation": f"์˜ค๋ฅ˜: {exc}",
            "question_type": question_type,
            "code": "",
            "sequence": [],
        }