ai-llm-patterns

RAG architecture pipelines retrieve relevant document chunks via vector search and pass them to an LLM to generate answers. You chunk documents, generate embeddings, and use similarity search to supply context for accurate question answering.

Structured output with Zod lets you define TypeScript schemas that validate LLM responses. You enforce schema compliance on the model's output, ensuring the returned data matches your expected types and application contracts.

Token cost management and prompt caching reduce Anthropic model expenses. By caching frequent prompt prefixes and selecting appropriate models for specific tasks, you optimize token usage and lower operational costs.

Anthropic SDK streaming sends model responses incrementally as they generate, enabling real-time UI updates. You process tokens as they arrive rather than waiting for the full response, improving perceived application latency.

Tool use and function calling allow LLMs to interact with external tools and APIs. You define available functions, and the model decides when to invoke them, bridging the AI model with your application's external capabilities.

Model selection involves matching LLM capabilities to your task complexity and latency needs. You evaluate factors like context window size, reasoning ability, and cost to pick the optimal model for your specific AI feature.