Using SDKs Directly

Before you reach for a heavyweight framework, you should know how to talk to AI providers with nothing but their official SDKs. Every major provider ships a thin client library — openai, @anthropic-ai/sdk, @google/genai — that wraps its HTTP API with typed methods, authentication, retries, and streaming helpers. Working at this level means you see exactly what goes over the wire: the model name, the messages, the parameters, and the raw response coming back.

This matters because abstraction layers hide the details you will eventually need to debug. When a framework mis-formats a tool call or silently swallows a rate-limit error, developers who only know the framework are stuck. If you understand the underlying SDK calls, you can read provider docs directly, reproduce problems with a ten-line script, and make informed decisions about cost, latency, and error handling. Direct SDK usage is also usually less code than the framework equivalent for simple pipelines.

In practice, you will install the provider package, load an API key from an environment variable, and compose small building blocks yourself: an embeddings call for search, a chat call for generation, a text splitter for chunking documents, and a vector store like Supabase’s pgvector for retrieval. You wire these together with plain functions, which keeps the whole pipeline inspectable and easy to test.