基于SambaNova的DeepSeek-R1、Qdrant二进制量化以及LangGraph，实现了32倍内存缩减的一个构建快速RAG系统方案

from qdrant_client import QdrantClient, models

client = QdrantClient(
    url="http://localhost:6333",
    prefer_grpc=True,
)

# 名称可自定义
collection_name = "binary-quantization"

if not client.collection_exists(collection_name):
    client.create_collection(
        collection_name=collection_name,
        vectors_config=models.VectorParams(
            size=1536,                   # 嵌入维度
            distance=models.Distance.DOT, # 点积距离
            on_disk=True,                # 原始向量存磁盘
        ),
        optimizers_config=models.OptimizersConfigDiff(
            default_segment_number=5,
        ),
        hnsw_config=models.HnswConfigDiff(
            m=0,                          # HNSW 参数（示例）
        ),
        quantization_config=models.BinaryQuantization(
            binary=models.BinaryQuantizationConfig(always_ram=True),
        ),
    )
else:
    print("Collection 已存在")

pip install langgraph langchain langchain-community \
            langchain-qdrant fastembed langchain-sambanova pypdf

from langchain_community.document_loaders import PyPDFDirectoryLoader

loader = PyPDFDirectoryLoader(
    path="./data/",
    glob="**/[!.]*.pdf",
    extract_images=False,
    mode="page",
    extraction_mode="plain",
)
docs = loader.load()
# 清理换行与制表符
for d in docs:
    d.page_content = d.page_content.replace("\n", " ").replace("\t", " ")

from langchain_community.embeddings import FastEmbedEmbeddings
from langchain_qdrant import QdrantVectorStore

embeddings = FastEmbedEmbeddings(model_name="thenlper/gte-large")
vector_store = QdrantVectorStore(
    client=client,
    collection_name=collection_name,
    embedding=embeddings,
)
vector_store.add_documents(docs)

from langchain_sambanova import ChatSambaNovaCloud
from langgraph.graph import START, StateGraph
from langchain_core.documents import Document
from typing_extensions import TypedDict

# 定义状态结构
class State(TypedDict):
    question: str
    context: list[Document]
    answer: str

# 初始化 LLM
import os
from google.colab import userdata
os.environ['SAMBANOVA_API_KEY'] = userdata.get("SAMBANOVA_API_KEY")
llm = ChatSambaNovaCloud(
    model="DeepSeek-R1",
    max_tokens=1024,
    temperature=0.1,
    top_p=0.01,
)

# 检索节点
def search(state: State):
    docs = vector_store.max_marginal_relevance_search(state["question"])
    return {"context": docs}

# 生成节点
SYSTEM_TEMPLATE = """…"""
HUMAN_TEMPLATE  = """…"""

def generate(state: State):
    ctx = "\n\n".join(
        f"{d.page_content}\nMetadata: Source-{d.metadata['source']} Page-{d.metadata['page_label']}"
        for d in state["context"]
    )
    messages = [
        {"role": "system", "content": SYSTEM_TEMPLATE},
        {"role": "user",   "content": HUMAN_TEMPLATE.format(context_str=ctx, query=state["question"])},
    ]
    resp = llm.invoke(messages)
    return {"answer": resp.content}

# 组装图
graph_builder = StateGraph(State).add_sequence([search, generate])
graph_builder.add_edge(START, "search")
graph = graph_builder.compile()

# 执行示例
response = graph.invoke({"question": "如何快速进入工作状态？"})
print(response["answer"])