프로덕션 아키텍처 설계

# Component inventory for a Document QA Assistant:
COMPONENTS = [
    'document_ingestion',   # PDF/Word -> chunks -> embeddings -> pgvector
    'hybrid_retriever',     # BM25 + dense + RRF
    'reranker',             # Cohere rerank
    'qa_agent',             # GPT-4o with RAG + function calling
    'semantic_cache',       # Redis + embedding similarity
    'streaming_api',        # FastAPI StreamingResponse
    'tracing',              # LangSmith or Langfuse
    'prompt_injection_filter', # Input sanitization
    'eval_pipeline',        # Automated quality scoring
]

# Technology decisions and their rationale
STACK = {
    'api':           ('FastAPI',    'Async support, OpenAPI docs, streaming easy'),
    'vector_store':  ('pgvector',   'Already on PostgreSQL, no extra infra'),
    'llm_primary':   ('gpt-4o',     'Best quality for the use case'),
    'llm_fallback':  ('claude-3.5', 'Different provider for resilience'),
    'cache':         ('Redis',      'Sub-ms lookup, TTL support built-in'),
    'tracing':       ('LangSmith',  'Native LangChain integration'),
    'embedding':     ('text-embedding-3-small', 'Good quality, low cost'),
    'reranker':      ('Cohere',     'Best-in-class rerank API'),
}

# Data flow (ASCII art):
#
# INGESTION PATH (offline):
# Documents --> Loader --> Chunker --> Embedder --> pgvector
#                                             |
#                                         BM25 index
#
# QUERY PATH (online):
# User Query
#    |-> Semantic Cache (hit: return) -> miss:
#    |-> Hybrid Retriever (BM25 + dense)
#    |-> Cohere Reranker
#    |-> LangChain LCEL Chain
#    |-> GPT-4o (streaming) --> FastAPI StreamingResponse
#    |-> LangSmith (trace every step)

from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI(title='Document QA Assistant', version='1.0')

class QueryRequest(BaseModel):
    question: str
    conversation_id: str | None = None
    max_chunks: int = 5
    stream: bool = True

class QueryResponse(BaseModel):
    answer: str
    sources: list
    cached: bool
    latency_ms: int
    trace_id: str

@app.post('/query', response_model=QueryResponse)
async def query_endpoint(request: QueryRequest):
    pass  # implementation next lesson

# Cost estimation model
DAU = 100           # daily active users
QPD = 10            # queries per user per day
TOKENS_PER_QUERY = {
    'prompt_tokens': 2000,  # system + context + question
    'completion_tokens': 500
}

PRICE_GPT4O_INPUT = 2.50 / 1_000_000   # per token
PRICE_GPT4O_OUTPUT = 10.00 / 1_000_000  # per token

daily_cost = DAU * QPD * (
    TOKENS_PER_QUERY['prompt_tokens'] * PRICE_GPT4O_INPUT +
    TOKENS_PER_QUERY['completion_tokens'] * PRICE_GPT4O_OUTPUT
)
print(f'Daily: ${daily_cost:.2f}, Monthly: ${daily_cost * 30:.2f}')

from dataclasses import dataclass
from typing import Optional

@dataclass
class ChunkMetadata:
    doc_id: str
    source_file: str
    page_number: Optional[int]
    section_title: Optional[str]
    created_at: str
    author: Optional[str]
    doc_type: str  # 'pdf', 'docx', 'html'

# Ingestion config
INGESTION_CONFIG = {
    'chunk_size': 800,
    'chunk_overlap': 100,
    'embedding_model': 'text-embedding-3-small',
    'embedding_dimensions': 1536,
    'batch_size': 100,  # chunks per embedding API call
}

SECURITY_DECISIONS = {
    'auth': 'JWT with 24h expiry',
    'data_isolation': 'tenant_id column in all vector metadata, filter on every query',
    'injection_detection': 'rule-based pre-filter + LLM secondary check for complex inputs',
    'output_scanning': 'check for PII, system prompt leakage patterns',
    'destructive_actions': 'require confirmation token for delete operations',
    'rate_limiting': '20 queries/minute per user, 429 with Retry-After header',
    'key_storage': 'AWS Secrets Manager, rotated every 90 days'
}

SUCCESS_METRICS = {
    # Quality
    'min_correctness_score': 4.0,       # out of 5, LLM-as-judge
    'min_retrieval_hit_rate': 0.85,     # top-5 chunk contains answer
    # Latency
    'p95_ttft_ms': 800,                 # time to first token
    'p95_total_latency_ms': 8000,       # full response
    # Cost
    'max_cost_per_query_usd': 0.05,     # $0.05 per Q&A
    # Reliability
    'target_uptime': 0.999,             # 99.9%
    'cache_hit_rate_target': 0.25,      # 25% queries served from cache
}

# Architecture Decision Records (ADRs):
#
# ADR-001: Use pgvector for vector storage
# Decision: pgvector in existing PostgreSQL
# Alternatives: Pinecone, Weaviate, Qdrant
# Reason: No new infra, row-level security native, familiar operations
# Trade-offs: Limited to ~5M vectors before performance degrades
#
# ADR-002: GPT-4o as primary model
# Decision: gpt-4o for all user-facing queries
# Alternatives: gpt-4o-mini (cheaper), Claude (alternative)
# Reason: Highest quality for use case, Claude as fallback
# Trade-offs: $0.04/query vs $0.002 for gpt-4o-mini

# Deployment topology:
#
# Internet -> Load Balancer (AWS ALB)
#                |
#            FastAPI API  (stateless, 2-4 containers, auto-scale)
#                |
#         +------+------+
#         |             |
#    pgvector        Redis Cache
#  (AWS RDS Postgres) (Elasticache)
#         |
#    Ingestion Worker  (separate container, manual trigger)
#         |
#    LangSmith (external SaaS, traces only)
#
# All containers in same VPC, no public access to DB/cache