AI CHATBOTS, ASSISTANTS AND SEARCH
Let visitors ask your website a question and get an answer from your own content, with sources. Give your editors an assistant that builds pages from a chat. All of it runs on Drupal, and you can try it today on our Drupal AI demo site.
Three ways a conversation beats a search box
- Ask AI on your website: visitors ask in their own words and get a short answer written from your pages only, with a citation on every claim. A follow-up question stays in the same conversation.
- An assistant for your editors: describe a landing page and an agent builds it from the components your site already has. Ask for a component that is missing and it writes one.
- Your content from the tools you already use: connect Claude or another AI client to your site over MCP and let it find and correct content, with Drupal's permissions and revisions in charge.
Each of these runs on aidemo.brainsum.com, our public Drupal AI demo site, with a step-by-step walkthrough per feature. Ask AI needs no login.
Ask AI: a cited answer instead of a list of links
The demo site is a fictional university with more than a hundred pages about programs, courses, departments and events. A visitor rarely knows which one holds the answer. Ask AI replaces the keyword search box with a question box: the site retrieves the pages that match, and an AI agent writes a short answer from those pages only. Every claim carries a numbered citation that links back to the page it came from.
- Follow-ups keep the context: "Which courses does it include?" is understood as the program asked about before.
- Every answer can be rated, which tells you which questions your content answers badly.
- Visitors can ask from the footer of any page, not only from the search page.
No invented answers
The answer is written from the retrieved page text, not from what the model knows about universities in general. Ask the demo how much a visitor parking permit costs and it tells you it has no price, then reports what the Visit Campus page does say about parking. If no page matches at all, the visitor gets a message you configure, not a guess.
The same retrieval works on an intranet, a support knowledge base or a set of PDF documents. The index lives in your own vector database, so the content never has to be uploaded to a third party's chatbot.
Canvas AI: pages and components built by chat
Drupal Canvas is Drupal's visual page builder, and it ships with two AI agents. Describe the page you need ("a landing page for the Marine Robotics Lab: a hero, three highlights of what students do, a call to action to book a tour") and the Page Builder agent answers with a plan, one entry per section, each mapped to a component your site already has. Say yes and it places the sections and writes the copy.
When a component is missing, the Component Agent writes a React component with editable properties, so editors can change the numbers and labels without touching code. Nothing is published until an editor reviews the changes, so an experiment costs nothing.
Your site as a tool for AI clients
With the Model Context Protocol (MCP), your Drupal site becomes a server that AI clients can talk to. On the demo, Claude.ai is added as a connector once and signs in through the site's own OAuth flow, as the editor's own account. From then on an editor can ask in plain words for an article they only vaguely remember, have Claude find it through the site's search index, and then ask for a correction: Claude reads the article, notices the same name in the linked person profile, and fixes both. Edits appear in each page's revision history like any other edit, and whatever the editor's role may not change, Claude cannot change either. Any MCP client that supports OAuth can connect the same way. See the MCP walkthrough.
How it works, and why it stays yours
- A regular Search API index on a vector database. Pages are split into chunks and stored as embeddings in Postgres with pgvector, Milvus, Pinecone, Azure AI Search or SQLite. The same fields, processors and cron indexing your site already uses. Ask AI and the MCP search tool both query this index.
- An agent with written instructions. The answering agent is told to search first, answer only from retrieved content and cite inline. The instructions, the no-answer message, the way sources are shown and the feedback settings are configuration, not code.
- Content tools published over MCP. The site exposes tools to search, read, create, edit and delete content. A client registers itself as an OAuth client and each user signs in with their own Drupal account; there is no shared API key. The client asks before using a tool, every write goes through Drupal's normal content handling and creates a revision, and permissions decide what a tool may touch.
- Any model, including your own. Drupal AI integrates 48 providers, among them OpenAI, Anthropic, Google Gemini, Mistral, AWS Bedrock and Azure, as well as local models through Ollama or LM Studio for data that must not leave your infrastructure. BRAINSUM contributed to the Anthropic (Claude) provider.
- Controls built in. Every AI request is logged, guardrails block prompts that contain personal data, and token usage and cost are metered per user and per role, inside Drupal. See AI Observability and AI Metering on the demo.
We started early. Now we help build Drupal AI.
In November 2024, shortly after the first release of the Drupal AI module, we built a RAG chatbot demo on a recipe site: Drupal, a Milvus vector database and OpenAI, with the chatbot's tone set by a prompt. It is still online.
Since then BRAINSUM has become a Certified Drupal AI Partner. We contribute to the Drupal AI Initiative's modules and recipes, and our demo site is built from the initiative's own starter kit. What was a chatbot in the corner of a page is now cited answers, page-building agents and MCP access, each shipped as a recipe that can be installed on your site.
Try the demo, then talk to us
Ask the demo site a question, build a page with Canvas AI, then tell us which of these you want on your own site. We start from your content and your business case, and we take care of the model choice, the hosting of your data and the controls around it.