Feature/us 09 resume analyzer - PDF 이력서 to Markdown 구현

ai/src/ai_server/analyzer/resume_analyzer.py

@@ -0,0 +1,88 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+import structlog
+
+from ai_server.analyzer.sources.base import SourceExtractor
+from ai_server.chain.document_analysis_chain import DocumentAnalyzer
+from ai_server.storage.base import ObjectStorage
+
+log = structlog.get_logger(__name__)
+
+
+class ResumeAnalyzeError(Exception):
+    def __init__(self, *, code: str, message: str, retriable: bool) -> None:
+        super().__init__(message)
+        self.code = code
+        self.message = message
+        self.retriable = retriable
+
+
+@dataclass(frozen=True)
+class ResumeAnalysisResult:
+    summary: str
+    tech_stack: list[str]
+    document_path: str
+
+
+# 스토리지에서 가져오고 LLM을 통해 분석함
+class ResumeAnalyzer:
+    def __init__(
+        self,
+        *,
+        extractor: SourceExtractor,
+        chain: DocumentAnalyzer,
+        storage: ObjectStorage,
+        analyzed_key_template: str,
+    ) -> None:
+        self._extractor = extractor
+        self._chain = chain
+        self._storage = storage
+        self._analyzed_key_template = analyzed_key_template
+
+    async def analyze(
+        self,
+        *,
+        resume_id: int,
+        file_path: str,
+    ) -> ResumeAnalysisResult:
+        log.info(
+            "resume.extract.start",
+            resume_id=resume_id,
+            file_path=file_path,
+        )
+        source = await self._extractor.extract(file_path)
+
+        if not source.text.strip():
+            raise ResumeAnalyzeError(
+                code="EMPTY_PDF_TEXT",
+                message="추출된 텍스트가 비어있음",
+                retriable=False,
+            )
+
+        log.info(
+            "resume.llm.start",
+            resume_id=resume_id,
+            text_chars=len(source.text),
+            source_type=source.source_type,
+        )
+        analysis = await self._chain.analyze(
+            text=source.text,
+            source_type=source.source_type,
+        )
+
+        out_key = self._analyzed_key_template.format(resume_id=resume_id)
+        await self._storage.put_text(out_key, analysis.markdown)
+        log.info(
+            "resume.markdown.saved",
+            resume_id=resume_id,
+            key=out_key,
+            md_chars=len(analysis.markdown),
+        )
+
+        return ResumeAnalysisResult(
+            summary=analysis.summary,
+            tech_stack=list(analysis.tech_stack),
+            document_path=out_key,
+        )

ai/src/ai_server/analyzer/sources/__init__.py

@@ -0,0 +1,13 @@
+from ai_server.analyzer.sources.base import (
+    ExtractedSource,
+    SourceExtractor,
+    SourceType,
+)
+from ai_server.analyzer.sources.pdf import PdfSourceExtractor
+
+__all__ = [
+    "ExtractedSource",
+    "SourceExtractor",
+    "SourceType",
+    "PdfSourceExtractor",
+]

ai/src/ai_server/analyzer/sources/base.py

@@ -0,0 +1,22 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+SourceType = Literal["PDF", "REPOSITORY", "WEB"]
+
+
+# 모든 Source Extractor가 공통으로 반환하는 결과 모델
+class ExtractedSource(BaseModel):
+    text: str
+    source_type: SourceType
+    metadata: dict[str, Any] = Field(default_factory=dict)
+
+
+# 단일 입력 형태에 대한 Adapter
+class SourceExtractor(ABC):
+
+    @abstractmethod
+    async def extract(self, locator: str) -> ExtractedSource: ...

ai/src/ai_server/analyzer/sources/pdf.py

@@ -0,0 +1,33 @@
+from __future__ import annotations
+
+import asyncio
+import io
+
+from pypdf import PdfReader
+
+from ai_server.analyzer.sources.base import ExtractedSource, SourceExtractor
+from ai_server.storage.base import ObjectStorage
+
+
+def _extract_pdf_text(data: bytes) -> str:
+    reader = PdfReader(io.BytesIO(data))
+    parts: list[str] = []
+    for page in reader.pages:
+        parts.append(page.extract_text() or "")
+    return "\n\n".join(p for p in parts if p).strip()
+
+
+# PDF 를 읽어 페이지 텍스트를 이어붙인다.
+class PdfSourceExtractor(SourceExtractor):
+
+    def __init__(self, storage: ObjectStorage) -> None:
+        self._storage = storage
+
+    async def extract(self, locator: str) -> ExtractedSource:
+        data = await self._storage.get_bytes(locator)
+        text = await asyncio.to_thread(_extract_pdf_text, data)
+        return ExtractedSource(
+            text=text,
+            source_type="PDF",
+            metadata={"locator": locator, "bytes": len(data)},
+        )

ai/src/ai_server/chain/__init__.py

@@ -0,0 +1,13 @@
+from ai_server.chain.document_analysis_chain import (
+    DocumentAnalysisResult,
+    DocumentAnalyzer,
+    LlmDocumentAnalyzer,
+    build_document_analysis_chain,
+)
+
+__all__ = [
+    "DocumentAnalysisResult",
+    "DocumentAnalyzer",
+    "LlmDocumentAnalyzer",
+    "build_document_analysis_chain",
+]

ai/src/ai_server/chain/document_analysis_chain.py

@@ -0,0 +1,62 @@
+from __future__ import annotations
+
+from typing import Protocol
+
+from langchain_core.output_parsers import PydanticOutputParser
+from langchain_core.prompts import ChatPromptTemplate
+from langchain_core.runnables import Runnable
+from pydantic import BaseModel, Field
+
+from ai_server.analyzer.sources.base import SourceType
+from ai_server.chain.prompts.document_analysis import HUMAN_PROMPT, SYSTEM_PROMPT
+from ai_server.config.settings import Settings
+
+
+class DocumentAnalysisResult(BaseModel):
+    summary: str = Field(..., description="2~4 sentence Korean summary")
+    tech_stack: list[str] = Field(default_factory=list)
+    markdown: str = Field(..., description="Interviewer-facing markdown")
+
+
+class DocumentAnalyzer(Protocol):
+    async def analyze(
+        self, *, text: str, source_type: SourceType
+    ) -> DocumentAnalysisResult: ...
+
+
+# 랭체인 파이프라인 호출을 감싼다
+class LlmDocumentAnalyzer:
+
+    def __init__(self, chain: Runnable) -> None:
+        self._chain = chain
+
+    async def analyze(
+        self, *, text: str, source_type: SourceType
+    ) -> DocumentAnalysisResult:
+        result = await self._chain.ainvoke({"text": text, "source_type": source_type})
+        if not isinstance(result, DocumentAnalysisResult):
+            raise TypeError(
+                f"chain returned {type(result).__name__}, expected DocumentAnalysisResult"
+            )
+        return result
+
+
+# 프롬프트 -> LLM -> 파서 하나로 묶어서 처리함
+def build_document_analysis_chain(settings: Settings) -> Runnable:
+    from langchain_openai import ChatOpenAI
+
+    parser = PydanticOutputParser(pydantic_object=DocumentAnalysisResult)
+    prompt = ChatPromptTemplate.from_messages(
+        [
+            ("system", SYSTEM_PROMPT),
+            ("human", HUMAN_PROMPT),
+        ]
+    ).partial(format_instructions=parser.get_format_instructions())
+
+    llm = ChatOpenAI(
+        model=settings.llm_pro_model,
+        temperature=settings.llm_pro_temperature,
+        api_key=settings.llm_api_key or None,
+        base_url=settings.llm_base_url,
+    )
+    return prompt | llm | parser

ai/src/ai_server/chain/prompts/document_analysis.py

@@ -0,0 +1,23 @@
+# 문서 분석 프롬프트 템플릿
+# source_type 으로 PDF, 웹, 리포지토리 구분함
+SYSTEM_PROMPT = (
+    "당신은 IT 직군 채용을 진행하는 시니어 면접관입니다. "
+    "지원자 자료(이력서·레포 README·기술 블로그 등)를 분석하여 면접 준비에 사용할 "
+    "구조화된 결과를 산출하세요.\n"
+    "- 사실에 근거해서 작성하고, 자료에 없는 내용을 추측해 채우지 마세요.\n"
+    "- 한국어로 작성하되 기술 용어는 영문 원어를 그대로 둡니다.\n"
+    "- 응답은 반드시 지정된 JSON 스키마를 따릅니다."
+)
+
+HUMAN_PROMPT = (
+    "다음은 지원자 자료입니다 (출처 유형: {source_type}).\n"
+    "---\n"
+    "{text}\n"
+    "---\n\n"
+    "요구 사항:\n"
+    "1. summary: 2~4문장 한국어 요약.\n"
+    "2. tech_stack: 자료에 명시된 핵심 기술 5~15개. 영문 표기.\n"
+    "3. markdown: 면접관이 훑어볼 한국어 마크다운. 섹션 구조는 "
+    "'## 개요', '## 주요 경험', '## 기술', '## 추가 메모' 사용.\n\n"
+    "{format_instructions}"
+)

ai/src/ai_server/config/settings.py

@@ -1,3 +1,5 @@
+from typing import Literal
+
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
@@ -14,25 +16,32 @@ class Settings(BaseSettings):
     debug: bool = False
 
     # RabbitMQ
-    rabbitmq_url: str = "amqp://stackup:stackup@localhost:5672/"
+    rabbitmq_url: str = "amqp://stackup:stackup@localhost:38050/"
 
-    # AI consumer
+    # AI consumer (큐 이름, prefetch, 콜백 라우팅 등)
     ai_queue_resume: str = "ai.analyze.resume"
     ai_queue_prefetch: int = 10
     ai_callback_exchange: str = "stackup.ai-to-core"
     ai_callback_routing_analysis: str = "callback.analysis"
     ai_publisher_name: str = "ai-server"
     ai_idempotency_lru_size: int = 1024
 
-    # S3 / MinIO
-    s3_endpoint_url: str = "http://localhost:9000"
+    storage_backend: Literal["local", "s3"] = "s3"
+    storage_local_root: str = "./var/storage"
+
+    s3_endpoint_url: str = "http://localhost:38060"
     s3_access_key: str = ""
     s3_secret_key: str = ""
     s3_bucket_name: str = "stackup"
+    s3_region: str = "us-east-1"
+
+    # 일단 충대 API 키 사용
+    llm_api_key: str = ""
+    llm_base_url: str = "https://factchat-cloud.mindlogic.ai/v1/gateway"
+    llm_pro_model: str = "gemini-3.1-pro-preview"
+    llm_pro_temperature: float = 0.2
 
-    # LLM API Keys
-    openai_api_key: str = ""
-    google_api_key: str = ""
+    analyzed_resume_md_key_template: str = "analyzed/resume/{resume_id}/summary.md"
 
 
 def get_settings() -> Settings: