On-site grounded search
Search that answers in sentences from your own pages, cites them, and says "I don't know" instead of guessing. Pilot scope, measured before rollout.
Answers come from your content only, each with a link to its page, and a plain "not on this site" when it is not there. We run it as a measured pilot in your own cloud account and roll it out only when the numbers hold up.
Talk about a search pilot- Answers cite your pages, never the open web
- Says "I don't know" instead of guessing
- Pilot scope, measured before rollout
What grounded search is for
Keyword search returns pages; people want answers. A grounded search takes the question, finds the passages on your site that bear on it, and writes an answer from those passages only, with a citation to each page. It is the useful half of a chatbot without the dangerous half: it cannot cite a page you did not publish, and when your content does not answer the question it declines rather than improvises. For a university, a hospital or an agency, that difference is the whole product.
What we do with it
We index your content as embeddings using the Drupal AI module's search, with the vector store and the model running inside your own cloud account. We build the answer interface into the site, with the citations visible and the decline path explicit. We set the boundaries: which content is in scope, what the system may never answer, and where a human contact is offered instead. And we measure it: a set of real questions with known answers, scored for right, wrong and correctly declined, before anyone outside the pilot sees it.
How we work
A pilot is two to four weeks on a slice of content, with named reviewers from your side and a question set drawn from real search logs and support tickets. We score every answer. If the wrong-answer rate is not low enough to defend, the pilot ends there and the report says why. If it is, rollout keeps the same measurement running monthly, and the cost is metered per query on your own cloud bill.
Why now
The Drupal AI module reached 1.4.8 in September 2026, with vector search and more than ninety-five providers behind one interface, so the pieces are standard rather than bespoke. Private endpoints in Bedrock, Azure OpenAI and Vertex keep the content in your account. This is a Horizon service, offered as a pilot only, and we say so here because the technology is good enough to try and not yet good enough to promise.
The rules of a pilot
1.4.8
Drupal AI module, current release
Released September 2, 2026, with vector search built in. Runs on Drupal 10.5 and 11.2 or later.
Drupal AI on drupal.org0
Answers from the open web
Every answer is written from your pages and cites them. Nothing else is in scope.
2 to 4
Weeks for a pilot
A slice of content, named reviewers, a real question set, every answer scored.
Yours
Cloud account, model and index
Bedrock, Azure OpenAI or Vertex, under your keys. Cost metered per query on your bill.
Our cloud engineering serviceI want to …
run a pilot on a slice of my content
Two to four weeks, one content area, real questions from your search logs and support tickets, every answer scored by your reviewers.
- A question set with known answers, drawn from real logs
- Right, wrong and correctly declined answers, counted
- A go or no-go report you can defend
answer staff or member questions from policy
The handbook, the policies, the procedures: an answer with a citation to the section, and a decline when the section does not exist.
- Scope limited to the documents you name
- Every answer links the section it came from
- Sensitive topics routed to a person, by rule
keep the whole thing inside my cloud account
The index, the model and the logs run in Bedrock, Azure OpenAI or Vertex under your keys. Nothing leaves your tenancy.
- Provider chosen for your cloud and your procurement rules
- Queries and answers logged in your account
- Cost per query metered on your own bill
make my content answerable in the first place
Grounded search is only as good as the pages behind it. Structured, headed, plainly written content answers well; a PDF scan does not.
- A content check: what is answerable, what is not, what to fix
- Headings, structure and plain language where the questions land
- The same work that makes the pages accessible and citable
A pilot, from question set to go or no-go
Every pilot follows the same gate. The automated steps run daily; the human ones decide.
A search that declines correctly is doing its job. The wrong-answer rate is the number the pilot exists to find.
Services that pair with this one
Grounded search is an AI feature on a content model in a cloud account. These are the services most often bought alongside.
-
AI engineering
We put language models to work inside Drupal on the jobs they do well: alt text, summaries, tagging, and answers from your own pages with citations. They run in your tenancy, your content trains no one else's model, and nothing publishes without a person saying so.
-
Cloud engineering
We design and run the cloud your platform lives on, in your own account, as code, with monitoring, patching and a monthly cost review in writing. We do not resell hosting, and everything we build is documented so the next engineer can run it.
-
Accessibility & AI visibility
Accessibility fixed at the source, in the theme and the editorial rules, with a conformance report you can hand to procurement. The same work makes your pages the ones AI engines cite.
Tell us the questions.
The ten questions people ask most, and where the answers live today. We come back with what a pilot would cover and what it would cost.
Talk about a search pilot