Build an AI chatbot with AI

OpenAI (GPT-4o / GPT-5). Default pick for general-purpose support and lead-gen bots. Cheaper per token, faster first-token latency, broadest model selection (4o-mini for cheap routing, 4o for quality, o1 for reasoning). The right default if you don't already have an opinion.

Anthropic Claude (Sonnet 4.6 / Opus 4.7). The right pick when conversation quality matters more than cost, sales chatbots, in-app product help, anything where tone and refusal behavior affect the experience. Claude is also stronger on long context (200K), so it shines for "talk to my docs" use cases where you stuff a giant knowledge base into the prompt.

Both, with routing. The mature pattern for serious chatbots: send short, simple questions to GPT-4o-mini at $0.15/M tokens, and route complex or high-stakes questions to Claude Sonnet at $3/M. Webtwizz scaffolds the routing layer if you ask for it, you describe the routing rule ("if message > 500 chars OR mentions account/billing, use Claude"), the AI writes the if/else.

Both integrations are pre-wired in Webtwizz. Connect once via OAuth, the keys are exposed as OPENAI_API_KEY and ANTHROPIC_API_KEY in your project's env vars, ready for the generated code to consume.

If your chatbot needs to answer questions grounded in your own data (docs, support articles, product specs), you need embeddings + vector search, not just a long prompt.

The pattern Webtwizz scaffolds: a Supabase table with a pgvector column, an embedding job that runs when content changes, and a retrieval step in the chat route that pulls the top-K most relevant chunks before calling the LLM. This is "RAG". Retrieval-augmented generation. It's how every serious doc chatbot is built.

For most B2B chatbots under 10K documents, pgvector inside Supabase is the right call: one fewer service to run, queries co-located with your auth and user data. Past that, you graduate to Pinecone or Qdrant. Webtwizz can wire either; pgvector is the default because it's simpler.

What you should NOT do: stuff your entire knowledge base into the system prompt every request. Even with Claude's 200K context, you'll burn money and quality. Retrieve, then prompt.