基于Python、LangChain、OpenAI的ChatBot无法生成NDA草稿问题排查

我正在构建一个基于Python、LangChain、OpenAI及LlamaIndex的ChatBot，目标是输入类似“为Mike LLC与Fantasty World生成加州保密协议草稿”的提示，结合docs目录中的保密协议示例生成文档草稿，但ChatBot始终返回“无法生成草稿”的回复。我尝试添加识别“draft”关键词的逻辑，直接调用OpenAI API生成草稿，但该方案未利用到已构建的向量索引，问题仍未解决。

相关代码 #

向量索引创建代码 #

import sys
import os
import openai
import constants
import gradio as gr
from langchain.chat_models import ChatOpenAI

from llama_index import SimpleDirectoryReader, GPTListIndex, GPTVectorStoreIndex, LLMPredictor, PromptHelper, load_index_from_storage

# Disable SSL certificate verification (for debugging purposes)
os.environ['REQUESTS_CA_BUNDLE'] = ''  # Set it to an empty string

os.environ["OPENAI_API_KEY"] = constants.APIKEY
openai.api_key = os.getenv("OPENAI_API_KEY")
print(os.getenv("OPENAI_API_KEY"))

def createVecorIndex(path):
    max_input = 4096
    tokens = 512
    chunk_size = 600
    max_chunk_overlap = 0.1

    prompt_helper = PromptHelper(max_input, tokens, max_chunk_overlap, chunk_size_limit=chunk_size)

    #define llm
    llmPredictor = LLMPredictor(llm=ChatOpenAI(temperature=.7, model_name='gpt-3.5-turbo', max_tokens=tokens))

    #load data
    docs = SimpleDirectoryReader(path).load_data()

    #create vector index
    vectorIndex = GPTVectorStoreIndex(docs, llmpredictor=llmPredictor, prompt_helper=prompt_helper)
    vectorIndex.storage_context.persist(persist_dir='vectorIndex.json')

    return vectorIndex

vectorIndex = createVecorIndex('docs')

首次查询逻辑 #

def chatbot(input_index):
    query_engine = vectorIndex.as_query_engine()
    response = query_engine.query(input_index)
    return response.response

gr.Interface(fn=chatbot, inputs="text", outputs="text", title="Super Awesome Chatbot").launch()

优化后的查询逻辑 #

def chatbot(input_index):
    query_engine = vectorIndex.as_query_engine()

    # If the "draft" clause is active:
    if "draft" in input_index.lower():
        # Query the vectorIndex for relevant information/context
        vector_response = query_engine.query(input_index).response
        print(vector_response)
        # Use vector_response as context to query the OpenAI API for a draft
        prompt = f"Based on the information: '{vector_response}', generate a draft for the input: {input_index}"
        
        response = openai.Completion.create(
            engine="text-davinci-002",
            prompt=prompt,
            max_tokens=512,
            temperature=0.2
        )
        
        openai_response = response.choices[0].text.strip()
        
        return openai_response

    # If "draft" clause isn't active, use just the vectorIndex response
    else:
        print('else clause')
        return query_engine.query(input_index).response

解决方案 #

1. 排查向量索引查询有效性 #

• 验证文档加载：在createVecorIndex函数中添加print(len(docs))，确认docs目录下的文件是否被正确读取。如果输出为0，检查路径是否正确、文件格式是否为LlamaIndex支持的类型（如txt、pdf，特殊格式需额外处理）。
• 调整文档拆分参数：当前chunk_size=600可能过小，导致上下文碎片化。尝试将chunk_size改为1000、max_chunk_overlap改为0.2，优化文档拆分逻辑，提升检索相关性。
• 单独测试索引查询：直接运行query_engine.query("保密协议核心条款")，查看返回结果是否包含docs中的示例内容，确认索引是否有效。

2. 优化草稿生成的Prompt与模型调用 #

当前Prompt未明确引导模型复用示例结构，且使用的text-davinci-002模型已停止更新，建议调整为：

if "draft" in input_index.lower():
    vector_response = query_engine.query(input_index).response
    # 明确要求模型遵循示例格式
    prompt = f"""
请参考以下保密协议示例内容：
{vector_response}
严格遵循示例的条款结构和格式，为{input_index}生成符合加州法律要求的保密协议草稿，需包含保密范围、双方义务、期限、违约责任等核心条款。
"""
    # 使用gpt-3.5-turbo模型，采用ChatCompletion接口
    response = openai.ChatCompletion.create(
        model="gpt-3.5-turbo",
        messages=[
            {"role": "system", "content": "你是专业的法律文档生成助手，擅长根据示例生成合规的法律草稿。"},
            {"role": "user", "content": prompt}
        ],
        max_tokens=1024,
        temperature=0.2
    )
    openai_response = response.choices[0].message.content.strip()
    return openai_response

3. 添加错误捕获与日志 #

在关键逻辑处添加异常捕获，定位具体问题：

def chatbot(input_index):
    query_engine = vectorIndex.as_query_engine()
    try:
        if "draft" in input_index.lower():
            vector_response = query_engine.query(input_index).response
            if not vector_response:
                return "未检索到相关保密协议示例，请检查docs目录内容"
            # 后续Prompt与API调用逻辑
        else:
            return query_engine.query(input_index).response
    except Exception as e:
        return f"生成失败：{str(e)}"

4. 复用已构建的向量索引 #

当前每次启动都会重新创建索引，改为优先加载已持久化的索引，避免重复构建：

from llama_index import StorageContext

def get_or_create_vector_index(path):
    persist_dir = 'vectorIndex.json'
    if os.path.exists(persist_dir):
        storage_context = StorageContext.from_defaults(persist_dir=persist_dir)
        return load_index_from_storage(storage_context)
    else:
        max_input = 4096
        tokens = 512
        chunk_size = 1000
        max_chunk_overlap = 0.2
        prompt_helper = PromptHelper(max_input, tokens, max_chunk_overlap, chunk_size_limit=chunk_size)
        llmPredictor = LLMPredictor(llm=ChatOpenAI(temperature=.7, model_name='gpt-3.5-turbo', max_tokens=tokens))
        docs = SimpleDirectoryReader(path).load_data()
        vectorIndex = GPTVectorStoreIndex(docs, llmpredictor=llmPredictor, prompt_helper=prompt_helper)
        vectorIndex.storage_context.persist(persist_dir=persist_dir)
        return vectorIndex

vectorIndex = get_or_create_vector_index('docs')

内容的提问来源于stack exchange，提问作者Mike Mann