File size: 6,718 Bytes
9381502
 
 
 
 
 
 
 
88c7275
9381502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88c7275
 
 
 
 
9381502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88c7275
9381502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88c7275
9381502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88c7275
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9381502
 
 
 
 
 
 
 
 
 
 
88c7275
 
 
 
 
 
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
from __future__ import annotations

from functools import lru_cache
import hashlib
import time

from cert_study_app.chains.study_assistant_chain import build_ollama_llm, build_study_assistant_chain
from cert_study_app.models import Question
from cert_study_app.services.official_docs_service import LEGACY_AZURE_DOCS_COLLECTION, OFFICIAL_DOCS_COLLECTION
from cert_study_app.services.vector_service import QuestionVectorStore


_ANSWER_CACHE: dict[str, tuple[float, dict]] = {}
ANSWER_CACHE_TTL_SECONDS = 600
ANSWER_PROMPT_VERSION = "ko-summary-v2"


@lru_cache(maxsize=8)
def cached_ollama_llm(model: str, base_url: str, temperature: float, num_predict: int):
    return build_ollama_llm(model=model, base_url=base_url, temperature=temperature, num_predict=num_predict)


@lru_cache(maxsize=8)
def cached_study_chain(model: str, base_url: str, temperature: float, num_predict: int):
    llm = cached_ollama_llm(model, base_url, temperature, num_predict)
    return build_study_assistant_chain(llm)


def _cache_key(
    question: str,
    model: str,
    base_url: str,
    k: int,
    source: str | None,
    max_context_chars: int,
    embedding_model: str,
) -> str:
    raw = "|".join(
        [ANSWER_PROMPT_VERSION, question, model, base_url, str(k), source or "", str(max_context_chars), embedding_model]
    )
    return hashlib.sha1(raw.encode("utf-8")).hexdigest()


class StudyAssistantService:
    def __init__(self, db, vector_store: QuestionVectorStore | None = None):
        self.db = db
        self.vector_store = vector_store or QuestionVectorStore()
        self.docs_vector_store = QuestionVectorStore(
            collection_name=OFFICIAL_DOCS_COLLECTION,
            embedding_model=self.vector_store.embedding_model,
        )
        self.legacy_docs_vector_store = QuestionVectorStore(
            collection_name=LEGACY_AZURE_DOCS_COLLECTION,
            embedding_model=self.vector_store.embedding_model,
        )

    def index_questions(self) -> int:
        questions = self.db.query(Question).order_by(Question.id.asc()).all()
        payloads = []
        for question in questions:
            payloads.append(
                {
                    "id": question.id,
                    "stem": question.stem,
                    "options": question.get_options(),
                    "answer": question.answer,
                    "explanation": question.explanation,
                    "category": question.category,
                    "subcategory": question.subcategory,
                    "source": question.source,
                }
            )
        return self.vector_store.upsert_questions(payloads)

    def ask(
        self,
        question: str,
        model: str = "qwen2.5:14b",
        base_url: str = "http://localhost:11434",
        k: int = 2,
        source: str | None = None,
        max_context_chars: int = 2200,
        use_cache: bool = True,
    ) -> dict:
        key = _cache_key(question, model, base_url, k, source, max_context_chars, self.vector_store.embedding_model)
        now = time.time()
        if use_cache and key in _ANSWER_CACHE:
            cached_at, cached = _ANSWER_CACHE[key]
            if now - cached_at <= ANSWER_CACHE_TTL_SECONDS:
                return {**cached, "cached": True}

        doc_results = self._search_docs(question, k=min(3, max(1, k)))
        question_results = self.vector_store.search(question, k=k, source=source)
        results = [*doc_results, *question_results]
        chain = cached_study_chain(model, base_url, 0.2, 320)
        answer = chain.invoke(
            {
                "question": question,
                "results": results,
                "max_context_chars": max_context_chars,
            }
        )
        payload = {
            "answer": answer,
            "sources": [
                {
                    "id": item.id,
                    "score": item.score,
                    "text": item.text,
                    "metadata": item.metadata or {},
                }
                for item in results
            ],
            "cached": False,
        }
        if use_cache:
            _ANSWER_CACHE[key] = (now, payload)
        return payload

    def ask_stream(
        self,
        question: str,
        model: str = "qwen2.5:14b",
        base_url: str = "http://localhost:11434",
        k: int = 2,
        source: str | None = None,
        max_context_chars: int = 2200,
        use_cache: bool = True,
    ) -> dict:
        key = _cache_key(question, model, base_url, k, source, max_context_chars, self.vector_store.embedding_model)
        now = time.time()
        if use_cache and key in _ANSWER_CACHE:
            cached_at, cached = _ANSWER_CACHE[key]
            if now - cached_at <= ANSWER_CACHE_TTL_SECONDS:
                return {**cached, "cached": True, "stream": None}

        doc_results = self._search_docs(question, k=min(3, max(1, k)))
        question_results = self.vector_store.search(question, k=k, source=source)
        results = [*doc_results, *question_results]
        chain = cached_study_chain(model, base_url, 0.2, 320)
        sources = [
            {
                "id": item.id,
                "score": item.score,
                "text": item.text,
                "metadata": item.metadata or {},
            }
            for item in results
        ]

        def stream_chunks():
            chunks = []
            try:
                for chunk in chain.stream(
                    {
                        "question": question,
                        "results": results,
                        "max_context_chars": max_context_chars,
                    }
                ):
                    text = str(chunk or "")
                    if not text:
                        continue
                    chunks.append(text)
                    yield text
            except Exception as exc:
                error_msg = f"\n\n*(Ollama 응답 오류: {exc})*"
                chunks.append(error_msg)
                yield error_msg
                return
            if use_cache:
                _ANSWER_CACHE[key] = (
                    time.time(),
                    {
                        "answer": "".join(chunks),
                        "sources": sources,
                        "cached": False,
                    },
                )

        return {"stream": stream_chunks(), "sources": sources, "cached": False}

    def _search_docs(self, question: str, k: int):
        official_results = self.docs_vector_store.search(question, k=k)
        if official_results:
            return official_results
        return self.legacy_docs_vector_store.search(question, k=k)