python3 -m venv .envs/agentic-ai
source .envs/agentic-ai/bin/activate
mkdir agentic-ai 
cd agentic-ai

pip install langchain langchain-huggingface langchain-community pypdf langchain_chroma
pip install chromadb smolagents python-dotenv gradio sentence-transformers
pip 'install smolagents[openai]'

ollama pull deepseek-r1:7b  
ollama pull qwen2.5:7b-instruct

$ cat deepseek-r1-7b-8k.Modelfile 
FROM deepseek-r1:7b 
PARAMETER num_ctx 8192

$ cat qwen2.5-7b-instruct-8k.Modelfile
FROM qwen2.5:7b-instruct  
PARAMETER num_ctx 8192

ollama create deepseek-r1-7b-8k -f ./deepseek-r1-7b-8k.Modelfile
ollama create qwen2.5:7b-instruct-8k -f ./qwen2.5-7b-instruct-8k.Modelfile

from langchain_community.document_loaders import DirectoryLoader, PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import Chroma
import os
import shutil

def load_and_process_pdfs(data_dir: str):
    """從目錄加載PDF并拆分為塊。"""
    loader = DirectoryLoader(
        data_dir,
        glob="**/*.pdf",
        loader_cls=PyPDFLoader
    )
    documents = loader.load()

    # 將文檔拆分為塊
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=1000,
        chunk_overlap=200,
        length_function=len,
    )
    chunks = text_splitter.split_documents(documents)
    return chunks

def create_vector_store(chunks, persist_dir: str):
    """創(chuàng)建并持久化Chroma向量存儲(chǔ)。"""
    # 如果存在現(xiàn)有向量存儲(chǔ)，則刪除
    if os.path.exists(persist_dir):
        print(f"正在從 {persist_dir} 刪除現(xiàn)有向量存儲(chǔ)")
        shutil.rmtree(persist_dir)

    # 初始化HuggingFace嵌入
    embedding_model = HuggingFaceEmbeddings(
        model_name="sentence-transformers/all-mpnet-base-v2",
        model_kwargs={'device': 'cpu'}
    )

    # 構(gòu)建并持久化新的Chroma向量存儲(chǔ)
    print("正在構(gòu)建并保存新的向量存儲(chǔ)...")
    vector_db = Chroma.from_documents(
        documents=chunks,
        embedding=embedding_model,
        persist_directory=persist_dir
    )
    return vector_db

def main():
    # 定義輸入數(shù)據(jù)和向量數(shù)據(jù)庫(kù)的路徑
    input_data_dir = os.path.join(os.path.dirname(__file__), "data")
    vector_db_dir = os.path.join(os.path.dirname(__file__), "chroma_db")

    # 加載PDF并拆分為較小的塊
    print("開始處理PDF...")
    document_chunks = load_and_process_pdfs(input_data_dir)
    print(f"從PDF生成了 {len(document_chunks)} 個(gè)文檔塊")

    # 構(gòu)建并存儲(chǔ)向量數(shù)據(jù)庫(kù)
    print("正在構(gòu)建向量存儲(chǔ)...")
    vector_db = create_vector_store(document_chunks, vector_db_dir)
    print(f"向量存儲(chǔ)已成功創(chuàng)建并保存到 {vector_db_dir}")

if __name__ == "__main__":
    main()

$ python ingest_pdfs.py  
開始處理PDF...
從PDF生成了312個(gè)文檔塊
正在構(gòu)建向量存儲(chǔ)...
向量存儲(chǔ)已成功創(chuàng)建并保存到./agentic-ai/chroma_db
$ ls chroma_db 
chroma.sqlite3                       d4c48cab-0467-4ca8-ad42-bafa5eb83040
$

# reasoning_model.py
from smolagents import OpenAIServerModel, CodeAgent
import os

# 定義本地模型名稱
reasoning_model_id = "deepseek-r1-7b-8k"# 使用DeepSeek進(jìn)行推理

def get_model(model_id):
    """返回一個(gè)Ollama模型。"""
    return OpenAIServerModel(
        model_id=model_id,
        api_base="http://localhost:11434/v1",  # Ollama API端點(diǎn)
        api_key="ollama"
    )

# 使用DeepSeek創(chuàng)建推理模型
reasoning_model = get_model(reasoning_model_id)

# 創(chuàng)建推理智能體
reasoner = CodeAgent(tools=[], model=reasoning_model, add_base_tools=False, max_steps=2)

# rag_tool.py
from langchain_chroma import Chroma
from langchain_huggingface import HuggingFaceEmbeddings
from reasoning_model import reasoner  # 從reasoning_model.py導(dǎo)入reasoner
import os

# 初始化向量存儲(chǔ)和嵌入
embeddings = HuggingFaceEmbeddings(
    model_name="sentence-transformers/all-mpnet-base-v2",
    model_kwargs={'device': 'cpu'}
)

# 向量數(shù)據(jù)庫(kù)目錄
db_dir = os.path.join(os.path.dirname(__file__), "chroma_db")

# 初始化向量存儲(chǔ)
vectordb = Chroma(persist_directory=db_dir, embedding_function=embeddings)


def rag_with_reasoner(user_query: str) -> str:
    """
    在向量數(shù)據(jù)庫(kù)中搜索相關(guān)上下文，并使用推理模型生成回復(fù)。

    參數(shù)：
        user_query：用戶的問題。
    """
    # 檢索相關(guān)文檔
    docs = vectordb.similarity_search(user_query, k=3)
    context = "\n\n".join(doc.page_content for doc in docs)

    # 為推理模型創(chuàng)建提示
    prompt = f"""根據(jù)以下上下文，簡(jiǎn)潔地回答用戶的問題。
    如果信息不足，請(qǐng)為RAG提出更好的查詢建議。

上下文：
{context}

問題：{user_query}

答案："""

    # 使用reasoner（DeepSeek-R1）生成回復(fù)
    response = reasoner.run(prompt, reset=False)
    return response

# primary_agent.py
from smolagents import OpenAIServerModel, ToolCallingAgent
from rag_tool import rag_with_reasoner  # 從rag_tool.py導(dǎo)入工具函數(shù)
import os

def get_model(model_id):
    """返回一個(gè)Ollama模型。"""
    return OpenAIServerModel(
        model_id=model_id,
        api_base="http://localhost:11434/v1",  # Ollama API端點(diǎn)
        api_key="ollama"
    )

# 加載工具模型（Qwen-2.5:7b）
primary_model = get_model("qwen2.5:7b-instruct-8k")

# 使用Qwen創(chuàng)建主智能體以生成工具回復(fù)
primary_agent = ToolCallingAgent(tools=[rag_with_reasoner], model=primary_model, add_base_tools=False, max_steps=3)

# 導(dǎo)出智能體以供應(yīng)用程序使用
def get_primary_agent():
    return primary_agent

# app.py
from primary_agent import get_primary_agent
from smolagents import GradioUI

# 從primary_agent.py獲取主智能體
primary_agent = get_primary_agent()

# 啟動(dòng)界面的主函數(shù)
def main():
    GradioUI(primary_agent).launch()

if __name__ == "__main__":
    main()

$ python app.py                                                                                                  
* 正在本地URL運(yùn)行：http://127.0.0.1:7860