用户 Query
    │
    ▼
QueryProcessor ─── 分词 / Filter 解析 / 同义词扩展 / 问题路由
    │
    ╱              ╲
Dense 检索           Sparse 检索
DashScope 1024d      jieba + BM25
ChromaDB HNSW        JSON 倒排索引
    ╲              ╱
     RRF(k=60) 融合
          │
     LLM Rerank 精排
          │
  AnswerGenerator ─── 上下文拼装 → LLM 生成答案
          │
  CitationVerifier ─── 逐句核查引用依据

pip install -e .

# 摄入文档
python -c "
from src.ingestion.pipeline import IngestionPipeline
from src.core.settings import load_settings
p = IngestionPipeline(load_settings())
p.run(file_path='data/documents/your-file.md')
"

# 查询
python scripts/query.py --query "你的问题" --collection default

# 评估
python scripts/run_evaluation.py --test-set tests/fixtures/semiconductor_golden.json

# Dashboard
streamlit run src/observability/dashboard/app.py

llm:       {provider: "deepseek", model: "deepseek-v4-pro"}
embedding: {provider: "dashscope", model: "text-embedding-v3"}  # 阿里云，OpenAI 兼容
rerank:    {enabled: true, provider: "llm"}
retrieval: {dense_top_k: 20, sparse_top_k: 20, fusion_top_k: 10, rrf_k: 60}

DEEPSEEK_API_KEY=sk-xxx
EMBEDDING_API_KEY=sk-xxx

src/
├── core/query_engine/   # Hybrid Search, Reranker, Query Processor, Query Cache
├── core/response/       # Answer Generator, Citation Verifier
├── core/trace/          # TraceContext, TraceCollector
├── ingestion/           # Pipeline, Chunk Refiner, Metadata Enricher
├── libs/                # LLM, Embedding, VectorStore, Reranker, Splitter, Evaluator
├── mcp_server/          # MCP Protocol Server + 3 Tools
└── observability/       # Dashboard, Evaluation