해외 멘토를 AI로 도플갱어 구축하는 중 - 앞으로 공부 많이 된다

## 기술 스택

- **Frontend**: Next.js 16.1.3, React 18+, TypeScript, Tailwind CSS v4, Shadcn UI
- **Backend**: Python 3.11+, FastAPI, Pydantic v2, FSD 아키텍처
- **Vector DB**: Qdrant (Docker)
- **LLM**: OpenAI API (gpt-5-nano)
- **Embeddings**: Voyage AI (voyage-4-nano, 384 dims)
- **Tools**: yt-dlp, youtube-transcript-api, uv

프로젝트 초기 실행 시 Front/Backend/Docker를 한 번에 실행할 수 있는 스크립트
- 명령어로 모두 실행 가능하게
- README에 사용법 작성

Chunk classification 프롬프트 개선
- 목표: Creator 스타일 모방 + 지식 베이스 구축
- 샘플: "세일즈 콜 첫 5분 노하우" 콘텐츠 분석

"""Chunk classification prompts and JSON schema.

Used by: src/features/chunk_classification/classifiers/llm_classifier.py
Purpose: Classify text chunks as KEEP or SKIP for vector database storage.
"""

SYSTEM_PROMPT = """You are classifying transcript chunks for a creator knowledge base.

## SKIP criteria (if ANY apply):
- Preview/intro content ("in this video we will...", "let me show you...")
- Generic advice available anywhere (e.g., "be confident", "practice makes perfect")
- Visual-dependent content (e.g., "click here", "as you can see")
- Pure self-introduction without unique insight
- Ads, sponsors, promotional content
- Filler: greetings, closings, transitions without substance
- Meta-commentary about the content structure

## KEEP criteria (ALL must apply):
1. Contains SPECIFIC, ACTIONABLE information (exact phrases, frameworks, numbers, steps)
2. Reflects creator's UNIQUE perspective or methodology
3. Could answer a direct question about this topic

## Examples:
SKIP: "hey guys welcome to today's training what we're going to go look at is how you can nail the first five minutes"
→ Preview content, no actionable information yet

SKIP: "Building rapport is important in sales calls"
→ Generic advice, not specific

KEEP: "I use 4 steps: rapport → confirm time → set agenda → transition to pain. For rapport, I ask 'whereabouts are you calling in from' - neutral and works globally."
→ Specific framework with exact phrase

KEEP: "I always say: 'normally the way these calls go is we learn about you, I share my background, then we see if there are mutual next steps' - this gives them control while you lead"
→ Exact script with reasoning

When in doubt, SKIP. Quality over quantity.
Respond in JSON format."""

USER_TEMPLATE = """Classify this transcript chunk.

Chunk:
{chunk_text}"""

# JSON Schema for OpenAI Structured Output
SCHEMA: dict = {
    "type": "object",
    "properties": {
        "decision": {
            "type": "string",
            "enum": ["KEEP", "SKIP"],
            "description": "Classification decision",
        },
        "reason": {
            "type": "string",
            "description": "Brief reason for the decision",
        },
    },
    "required": ["decision", "reason"],
    "additionalProperties": False,
}