ai-feature-architect

Create a route handler that calls streamText with your model and messages, then return result.toDataStreamResponse(). On the client, use the useChat hook from ai/react to manage streaming state, input handling, and errors.

Chunk documents, embed them in batch with embedMany, and store vectors in a PostgreSQL pgvector column. At query time, embed the user query with embed, run a cosine similarity search, and pass the top chunks as context to streamText.

Use generateObject with a Zod schema defining your expected shape. The SDK validates the model output against the schema and returns a fully typed object, eliminating manual parsing and unsafe type casts.

Yes, tool calling works with providers like Anthropic and OpenAI through a unified tools API. Define each tool with a Zod parameters schema and an execute function, and set maxSteps to cap the number of tool-call rounds.

Set maxTokens on every streamText or generateText call, add per-IP and per-user rate limiting, and log token usage in the onFinish callback. For tool calling, set maxSteps to prevent infinite tool-call loops.

Use multi-model routing when task complexity varies: route simple requests to a cheap fast model like gpt-4o-mini and complex reasoning to a larger model. This balances latency and cost against output quality per request.