prompt-engineering-patterns

Define a Pydantic model describing the expected fields, then use LangChain's with_structured_output method or instruct the model to respond with JSON matching the schema. Parse the response and validate it against the model to catch malformed outputs.

Use SemanticSimilarityExampleSelector from LangChain with an embedding model like VoyageAI and a vector store such as Chroma. It retrieves the k most similar examples to each query and injects them into the prompt automatically.

Chain-of-thought prompting asks the model to reason step by step before answering, improving accuracy on multi-step problems. Use it for math, logic, or analysis tasks; simple factual queries usually do not benefit from the extra tokens.

Yes, Anthropic's prompt caching lets you mark long system prompts with cache_control so repeated requests reuse the cached prefix. This reduces both latency and token costs when the same system prompt is sent across many calls.

Models can produce invalid JSON due to ambiguous instructions, missing schema details, or token limits truncating output. Handle this with try/except around json.loads and Pydantic validation, plus a fallback prompt that extracts a simpler answer.