
Who this is for
- Developers who do text classification and information extraction
- Technical bloggers who want to improve the stability of Prompt
What problem does this article solve?
Too few examples, too biased, or inconsistent formats, lead to model learning error output patterns.
Key concepts
- few-shot
- Sample coverage
- Counterexample
- format consistency
- border sample
Walkthrough
- Design Task Definition
- Prepare positive examples and edge cases
- Fixed output format
- Compare Zero-Shot and FEW-Shot effects
Python code example
The companion demo is generated in:
cd examples/llm-guides/006_prompt_few_shot
python demo.py
The example follows two principles:
- From
OPENAI_API_KEY, load the key at runtime to avoid hard-coding it. - Without an API key, the example runs in mock mode so you can follow the workflow.
The essential call looks like this:
import os
api_key = os.getenv("OPENAI_API_KEY")
if not api_key:
print("未配置 OPENAI_API_KEY,进入 mock 模式。")
else:
# 真实项目中在这里初始化 SDK 客户端并发起请求。
print("已读取 OPENAI_API_KEY,可以调用模型。")
Common pitfalls
- Publishing after a single successful demo run leaves out error handling, logging and human review.
- Including API keys, real file paths, account details or business data in published articles or code.
- Readers can struggle to reproduce an example if it is unclear which parts are mocked and which run in production.
- Citing an API or framework feature without checking its documentation can mislead readers when versions change.
Summary
- LLM Guide 06 explains how to write few-shot examples that give the model the right pattern to follow. Too few examples, an unrepresentative selection, or inconsistent formatting can teach it the wrong output pattern.
- Make the workflow reproducible, reversible and easy to review. One good generation tells you little about how it will hold up in use.
- Before publishing, check current documentation for model capabilities, API parameters and third-party frameworks.
CSDN tags
Prompt, feed-shot, large model, Python, NLP
WeChat article summary
A reproducible experiment shows how a few representative examples can guide model behavior. Simply adding more examples misses the point.
Alternative title for Zhihu
- Why does Few-shot Prompt sometimes get worse?
- How many examples should be placed?
- How to design high-quality examples for large models?
References and further reading
- Check the OpenAI documentation for your model, SDK and account permissions.
- Python documentation on environment variables, virtual environments and exception handling.
- Consult the documentation for the tools used in the example, such as LangChain, LlamaIndex, Haystack or the vector database.