商业实战使用DeepSeek-R1构建本地RAG系统的完整方案02-优雅草卓伊凡

graph TD
    A[归巢网数据] --> B[DeepSeek-R1 Embedding]
    B --> C[向量数据库]
    D[用户提问] --> B
    D --> E[DeepSeek-R1 LLM]
    C --> E
    E --> F[智能回答]

# 基础环境
conda create -n deepseek python=3.10
conda activate deepseek
pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cu118

# DeepSeek-R1相关
pip install deepseek-ai
git clone https://github.com/deepseek-ai/DeepSeek-R1

from deepseek.data_processor import WebLoader, Chunker

# 初始化网页加载器
loader = WebLoader(
    base_url="https://www.guichao.com",
    max_depth=3,
    obey_robots=True
)



# 获取并清洗数据
documents = loader.load()
chunks = Chunker(
    chunk_size=512,
    overlap=50
).process(documents)

from deepseek.encoder import R1Embedder
from deepseek.vectorstore import R1VectorStore


![](https://doc2.youyacao.com/server/index.php?s=/api/attachment/visitFile&sign=1501c7e5b54dc150945e6dada786c0e6)
# 初始化嵌入模型
embedder = R1Embedder(model_path="local/path/to/r1-embedding")

# 生成向量
embeddings = embedder.encode([chunk.text for chunk in chunks])

# 创建向量库
vector_db = R1VectorStore(dimension=1024)  # DeepSeek-R1嵌入维度
vector_db.add_documents(chunks, embeddings)

# 保存索引
vector_db.save("guichao_index")

from deepseek.llm import R1Generator

# 初始化生成模型
generator = R1Generator(
    model_path="local/path/to/r1-generator",
    device="cuda:0"
)

# RAG流程
def ask_question(question):
    # 检索
    query_embedding = embedder.encode([question])[0]
    results = vector_db.search(query_embedding, top_k=3)

    # 生成
    context = "\n\n".join([doc.text for doc in results])
    prompt = f"""基于以下归巢网内容回答问题：
    {context}

    问题：{question}
    """

    answer = generator.generate(prompt, max_length=300)
    return answer, [doc.metadata for doc in results]

def hybrid_retrieve(query):
    # 向量检索
    vector_results = vector_db.search(query)
    # 关键词检索
    keyword_results = bm25_retriever.search(query)
    # 结果融合
    return reciprocal_rank_fusion(vector_results, keyword_results)

from functools import lru_cache

@lru_cache(maxsize=1000)
def cached_embedding(text):
    return embedder.encode([text])[0]

for token in generator.stream_generate(prompt):
    print(token, end="", flush=True)

# Dockerfile示例
FROM nvidia/cuda:12.1-base
RUN apt-get update && apt-get install -y python3.10
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY . .
CMD ["python", "api_server.py"]

# prometheus.yml
scrape_configs:
  - job_name: 'deepseek'
    static_configs:
      - targets: ['localhost:8000']

deepseek-tune --model r1-base \
             --data guichao_dataset.json \
             --lora_rank 8 \
             --output_dir tuned_model

from deepseek.quantization import quantize_model
quantize_model("r1-generator", output="r1-int4", bits=4)

# 同时处理多个请求
answers = generator.batch_generate(prompts, batch_size=4)