RAG-based support bot.
A support chatbot is only useful if it answers from your docs, not from whatever a general model half-remembers. The hosted "train a bot on your site" widgets get you there, then charge per message and keep your knowledge base on their servers behind their branding. If your documentation is already written down, the hard part is done, what's left is plumbing. This blueprint builds it: a retrieval chatbot that reads your Markdown and answers out of it, as code you own.
A customer support chatbot prototype wired with the Vercel AI SDK and Pinecone. It takes a Markdown documentation file, splits it into chunks, embeds them, and stores the vectors in Pinecone, so the bot retrieves the passages relevant to a question before answering. That's the RAG pattern: the model responds using what your docs actually say instead of guessing, which keeps answers grounded and on-topic. The result is a working chat interface that gives sourced answers.
You watch it come together in the task workspace: the agent scaffolds the project, wires the embedding and retrieval flow, and runs a test question to confirm the bot answers from your file before handing you a live preview. The code lands on a branch in your repository with a pull request ready.
The first version is a prototype you steer with plain follow-up messages:
Each follow-up wakes the same repository and the agent continues on its own code.
Do I need my own API keys? Yes, it runs on your own model and Pinecone keys, so the bot and its index stay under your account.
What if my docs change? Ask the agent to add a re-index step and updating the bot's knowledge becomes a single command.
Is one Markdown file the limit? No, that's just the starting point; tell the agent to ingest more files or formats and it expands the pipeline.