\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260124144905-6974dbe194ad4.jpg\")
<\/p>\n\u5927\u8bed\u8a00\u6a21\u578b\u53d1\u5c55\u5230\u4eca\u5929，\u5199\u63d0\u793a\u8bcd\u4e0d\u4ec5\u662f\u4e2a\u4f53\u529b\u6d3b，\u8fd8\u662f\u4e00\u4e2a\u6280\u672f\u6d3b\u3002\u63d0\u793a\u8bcd\u7684\u63aa\u8f9e\u7a0d\u5fae\u6539\u4e00\u4e0b，\u6027\u80fd\u6ce2\u52a8 20-50% \u662f\u5e38\u6709\u7684\u4e8b\u3002\u90a3\u80fd\u4e0d\u80fd\u8ba9 LLM \u81ea\u5df1\u6539\u8fdb\u63d0\u793a\u8bcd\u5462？<\/p>\n
\u672c\u6587\u4e3b\u8981\u8bb2\u7684\u5c31\u662f\u8fd9\u4e2a，\u4e00\u4e2a\u5904\u7406\u590d\u6742\u63a8\u7406\u4efb\u52a1\u7684 Agent，\u7ecf\u8fc7\u51e0\u8f6e\u81ea\u52a8\u8fed\u4ee3，\u51c6\u786e\u7387\u4ece 34% \u98d9\u5230 97%\u76f8\u5bf9\u63d0\u5347 184%\u3002\u4e0b\u9762\u662f\u5b8c\u6574\u7684\u6b65\u9aa4：\u4ece\u73af\u5883\u642d\u5efa\u5230\u4ee3\u7801\u5b9e\u73b0，\u8fd8\u5305\u542b\u4e86\u4f18\u7f3a\u70b9\u548c\u9002\u7528\u573a\u666f\u3002<\/p>\n
\u81ea\u52a8\u63d0\u793a\u8bcd\u4f18\u5316\u6709\u65f6\u4e5f\u53eb Automatic Prompt Engineering（APE），\u6838\u5fc3\u601d\u8def\u662f\u628a\u5199\u63d0\u793a\u8bcd\u5f53\u6210\u4f18\u5316\u95ee\u9898\u6765\u89e3\u3002\u6d41\u7a0b\u5927\u6982\u662f\u8fd9\u6837：\u5148\u62ff\u4e00\u4e2a\u521d\u59cb\u63d0\u793a\u8bcd（\u54ea\u6015\u5199\u5f97\u5f88\u4e00\u822c），\u5728\u4e00\u6279\u6807\u6ce8\u597d\u7684\u6837\u672c\u4e0a\u8dd1\u4e00\u904d，\u7528\u8bc4\u5206\u51fd\u6570（\u7cbe\u786e\u5339\u914d\u3001BLEU\u3001LLM-as-judge \u4e4b\u7c7b\u7684）\u7b97\u4e2a\u5206，\u7136\u540e\u8ba9 LLM \u751f\u6210\u65b0\u7684\u5019\u9009\u63d0\u793a\u8bcd，\u6311\u8868\u73b0\u597d\u7684\u7ee7\u7eed\u8fed\u4ee3，\u76f4\u5230\u5206\u6570\u4e0d\u518d\u6da8\u4e3a\u6b62\u3002<\/p>\n
2023 \u5e74\u7684\u8bba\u6587 \u201cAutomatic Prompt Optimization with Gradient-Free Methods\u201d \u628a\u8fd9\u5957\u65b9\u6cd5\u8bb2\u5f97\u6bd4\u8f83\u7cfb\u7edf\u3002\u4ece\u5de5\u5177\u5c42\u9762\u5c06DSPy\u3001LangChain \u7684 PromptOptimizer \u90fd\u80fd\u5e72\u8fd9\u4e2a\u6d3b，\u4e0d\u8fc7\u6709\u4e00\u4e2a\u4e13\u95e8\u7684\u5e93\u53eb OPIK \u3002<\/p>\n
OPIK \u8fd9\u5957\u4e1c\u897f\u6bd4\u8f83\u5438\u5f15\u4eba\u7684\u662f\u5b8c\u5168\u5f00\u6e90，Apache 2.0 \u534f\u8bae，\u5546\u7528\u4e5f\u4e0d\u82b1\u94b1；\u53ef\u4ee5\u7eaf\u672c\u5730\u8dd1，Ollama\u3001LM Studio\u3001\u4efb\u4f55 OpenAI \u517c\u5bb9\u7684\u63a5\u53e3\u90fd\u884c；\u81ea\u5e26\u8ffd\u8e2a\u4eea\u8868\u677f，\u529f\u80fd\u7c7b\u4f3c LangSmith \u4f46\u662f\u786e\u662f\u514d\u8d39；API \u8bbe\u8ba1\u7b80\u5355，\u4e00\u4e2a\u7c7b\u5c31\u80fd\u641e\u5b9a\u4f18\u5316\u6d41\u7a0b；\u8bc4\u4f30\u6307\u6807\u968f\u4fbf\u4f60\u5b9a\u4e49，\u53ea\u8981\u80fd\u5199\u6210\u4ee3\u7801\u5c31\u884c\u3002<\/p>\n
\u521b\u5efa\u865a\u62df\u73af\u5883<\/p>\n
python -m venv opik-apo
\n source opik-apo\/bin\/activate # Windows: opik-apo\\\\Scripts\\\\activate
\n pip install opik[all] datasets sentence-transformers<\/p>\n
\u672c\u5730 LLM \u5b89\u88c5（\u53ef\u9009\u4f46\u63a8\u8350）<\/p>\n
# Example with Ollama
\n ollama pull llama3.2:8b-instruct-qat
\n ollama serve<\/p>\n
\u51c6\u5907\u8bc4\u4f30\u6570\u636e\u96c6<\/p>\n
\u4e0d\u9700\u8981\u592a\u591a\u6570\u636e，20 \u5230 100 \u6761\u9ad8\u8d28\u91cf\u6837\u672c\u5c31\u591f\u4e86\u3002\u4e0b\u9762\u662f\u4e2a GSM8K \u98ce\u683c\u7684 JSONL \u4f8b\u5b50，\u4fdd\u5b58\u6210<\/p>\n
eval_data.jsonl<\/p>\n
：<\/p>\n
{"question": "Janet's ducks lay 16 eggs per day. She eats 3 for breakfast, gives 8 to friends, and her bakery uses the rest. How many eggs does the bakery receive?", "answer": "5"}
\n {"question": "A store sells apples for $2 each or 6 for $10. How much do you save per apple by buying 6?", "answer": "0.33"}<\/p>\n
\u5b8c\u6574\u4ee3\u7801<\/p>\n
import opik
\nfrom opik import track
\nfrom opik.opimization import MetaPromptOptimizer
\nfrom opik.opimization.evaluation_metrics import LevenshteinRatio
\nfrom opik.integrations.langchain import OpikTracer
\nfrom langchain_community.llms import Ollama
\nfrom langchain.prompts import PromptTemplate
\nimport json <\/p>\n
# 1. \u521d\u59cb\u5316 Opik（\u9ed8\u8ba4\u672c\u5730）
\nopik_client = opik.Opik() <\/p>\n
# 2. \u52a0\u8f7d\u8bc4\u4f30\u6570\u636e\u96c6
\ndef load_dataset(path="eval_data.jsonl"):
\n data = []
\n with open(path) as f:
\n for line in f:
\n item = json.loads(line)
\n data.append({"input": item["question"], "expected": item["answer"]})
\n return data <\/p>\n
eval_dataset = load_dataset() <\/p>\n
# 3. \u5b9a\u4e49\u4f60\u7684\u57fa\u7840\u63d0\u793a\u8bcd\u6a21\u677f
\ninitial_prompt = """You are an expert assistant. Your task is answer questions accurately and concisely. Consider the context carefully before responding. <\/p>\n
Question: {input}
\nAnswer:""" <\/p>\n
prompt_template = PromptTemplate.from_template(initial_prompt) <\/p>\n
# 4. \u521b\u5efa LLM（\u672c\u5730\u6216 OpenAI）
\nllm = Ollama(model="llama3.2:8b-instruct-qat", temperature=0.2)
\n# llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.2) # for cloud <\/p>\n
# 5. \u5b9a\u4e49\u4f60\u60f3\u8981\u4f18\u5316\u7684\u94fe
\n@track
\ndef math_chain(question):
\n prompt = prompt_template.format(input=question)
\n response = llm.invoke(prompt)
\n return response.strip() <\/p>\n
# 6. \u5b9a\u4e49\u8bc4\u4f30\u51fd\u6570
\ndef evaluation_function(task):
\n expected = task["expected"]
\n generated = math_chain(task["input"])
\n score = LevenshteinRatio()(expected, generated)
\n return score <\/p>\n
# 7. \u8fd0\u884c\u4f18\u5316\u5668
\noptimizer = MetaPromptOptimizer(
\n task="You are an expert at solving math word problems.",
\n evaluation_dataset=eval_dataset,
\n evaluation_function=evaluation_function,
\n llm=llm,
\n n_iterations=5, # \u901a\u5e38 3-7 \u6b21\u5c31\u8db3\u591f\u4e86
\n population_size=10
\n) <\/p>\n
best_prompt = optimizer.run()
\nprint("\\\\n=== BEST PROMPT FOUND ===\\\\n") <\/p>\n
print(best_prompt)<\/p>\n
\u8fd0\u884c\u811a\u672c：<\/p>\n
python optimize_prompt.py<\/p>\n
\u8dd1\u5b8c 5 \u8f6e\u8fed\u4ee3\u540e，\u6253\u5f00 http:\/\/localhost:5173 \u5c31\u80fd\u770b\u5230\u4eea\u8868\u677f，\u6bcf\u4e2a\u5019\u9009\u63d0\u793a\u8bcd\u7684\u5f97\u5206\u3001\u6700\u7ec8\u9009\u4e2d\u7684\u7248\u672c\u90fd\u5728\u4e0a\u9762\u3002<\/p>\n
\u4f18\u5316\u524d\u7684\u63d0\u793a\u8bcd（34% \u7cbe\u786e\u5339\u914d）：<\/p>\n
You are an expert assistant. Your task is answer questions accurately and concisely. Consider the context carefully before responding.<\/p>\n
\u4f18\u5316\u540e（97% \u7cbe\u786e\u5339\u914d）：
\n<\/p>\n
\u4f18\u70b9
\n![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/01\/20260124144905-6974dbe1b82b8.png\")
<\/p>\n
\u6027\u80fd\u63d0\u5347\u5e45\u5ea6\u53ef\u89c2，50-200% \u7684\u76f8\u5bf9\u6539\u8fdb\u5e76\u4e0d\u7f55\u89c1\u3002\u6574\u4e2a\u8fc7\u7a0b\u5b8c\u5168\u53ef\u590d\u73b0，\u6bcf\u4e00\u8f6e\u8fed\u4ee3\u90fd\u6709\u65e5\u5fd7\u8bb0\u5f55，\u65b9\u4fbf\u5ba1\u8ba1\u3002\u672c\u5730\u8fd0\u884c\u610f\u5473\u7740\u521d\u59cb\u914d\u7f6e\u4e4b\u540e\u5c31\u4e0d\u518d\u4ea7\u751f Token \u8d39\u7528\u3002\u4e5f\u4e0d\u4f9d\u8d56\u68af\u5ea6\u4efb\u4f55\u9ed1\u76d2 LLM \u90fd\u80fd\u7528\u3002\u8fc1\u79fb\u5230\u65b0\u4efb\u52a1\u4e5f\u7b80\u5355，\u6539\u51e0\u4e2a\u53c2\u6570\u5c31\u884c\u3002<\/p>\n
\u5c40\u9650\u6027<\/p>\n
\u8bc4\u4f30\u6570\u636e\u96c6\u7684\u8d28\u91cf\u81f3\u5173\u91cd\u8981：\u5982\u679c\u6570\u636e\u4e0d\u884c，\u4f18\u5316\u51fa\u6765\u7684\u63d0\u793a\u8bcd\u4e5f\u4e0d\u4f1a\u597d\u5230\u54ea\u53bb，\u5c31\u662f\u5178\u578b\u7684 garbage in, garbage out；\u672c\u5730 LLM \u751f\u6210\u5019\u9009\u63d0\u793a\u8bcd\u7684\u901f\u5ea6\u53ef\u80fd\u504f\u6162；\u5c0f\u6570\u636e\u96c6\u4e0a\u5bb9\u6613\u8fc7\u62df\u5408\u6a21\u578b\u53ef\u80fd\u53ea\u662f"\u8bb0\u4f4f"\u4e86\u7b54\u6848\u800c\u975e\u771f\u6b63\u5b66\u4f1a\u89e3\u9898；\u4f18\u5316\u51fa\u6765\u7684\u63d0\u793a\u8bcd\u6709\u65f6\u5019\u53c8\u957f\u53c8\u602a，\u867d\u7136\u6548\u679c\u597d，\u4f46\u53ef\u8bfb\u6027\u582a\u5fe7\u3002<\/p>\n
OPIK \u7684 MetaPromptOptimizer \u8ba9\u81ea\u52a8\u63d0\u793a\u8bcd\u53d8\u5f97\u5f88\u7b80\u5355：\u62ff\u4e00\u4e2a\u52c9\u5f3a\u80fd\u7528\u7684\u63d0\u793a\u8bcd，\u534a\u5c0f\u65f6\u5185\u5c31\u80fd\u628a\u5b83\u8c03\u6210\u63a5\u8fd1 SOTA \u6c34\u5e73\u7684 one-shot Agent，\u7701\u4e0b\u539f\u672c\u8981\u82b1\u51e0\u5468\u624b\u52a8\u8fed\u4ee3\u7684\u65f6\u95f4\u3002<\/p>\n
\u627e\u4e00\u4e2a\u73b0\u6709\u7684 Agent，\u6512 30 \u6761\u6807\u6ce8\u6837\u672c，\u8dd1\u4e00\u904d\u4e0a\u9762\u7684\u811a\u672c，\u770b\u51c6\u786e\u7387\u66f2\u7ebf\u5f80\u4e0a\u8d70\u5c31\u5bf9\u4e86\u3002\u63d0\u793a\u8bcd\u5de5\u7a0b\u7684\u4e0b\u4e00\u6b65，\u4e0d\u662f\u4eba\u5199\u51fa\u66f4\u597d\u7684\u63d0\u793a\u8bcd，\u800c\u662f\u8ba9\u7cfb\u7edf\u66ff\u4eba\u5199\u3002<\/p>\n
\u5730\u5740<\/p>\n
https:\/\/avoid.overfit.cn\/post\/be1860b7176b474b9389d88501402048<\/p>\n
\u4f5c\u8005：Minervee<\/p>\n","protected":false},"excerpt":{"rendered":"
\u5927\u8bed\u8a00\u6a21\u578b\u53d1\u5c55\u5230\u4eca\u5929，\u5199\u63d0\u793a\u8bcd\u4e0d\u4ec5\u662f\u4e2a\u4f53\u529b\u6d3b，\u8fd8\u662f\u4e00\u4e2a\u6280\u672f\u6d3b\u3002\u63d0\u793a\u8bcd\u7684\u63aa\u8f9e\u7a0d\u5fae\u6539\u4e00\u4e0b，\u6027\u80fd\u6ce2\u52a8 20-50% \u662f\u5e38\u6709\u7684\u4e8b\u3002\u90a3\u80fd\u4e0d\u80fd\u8ba9 LLM \u81ea\u5df1\u6539\u8fdb\u63d0\u793a\u8bcd\u5462？
\n\u672c\u6587\u4e3b\u8981\u8bb2\u7684\u5c31\u662f\u8fd9\u4e2a，\u4e00\u4e2a\u5904\u7406\u590d\u6742\u63a8\u7406\u4efb\u52a1\u7684 Agent，\u7ecf\u8fc7\u51e0\u8f6e\u81ea\u52a8\u8fed\u4ee3，\u51c6\u786e\u7387\u4ece 34% \u98d9\u5230 97%\u76f8\u5bf9\u63d0\u5347 184%\u3002\u4e0b\u9762\u662f\u5b8c\u6574\u7684\u6b65\u9aa4：\u4ece\u73af\u5883\u642d\u5efa\u5230\u4ee3\u7801\u5b9e\u73b0，\u8fd8\u5305\u542b\u4e86\u4f18\u7f3a\u70b9\u548c\u9002\u7528\u573a\u666f\u3002<\/p>\n","protected":false},"author":2,"featured_media":65401,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[6904,50,956,86],"topic":[],"class_list":["post-65403","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-6904","tag-50","tag-956","tag-86"],"yoast_head":"\n