AI Answers

AI Answers turns an existing AI Agent into a RAG-powered search answer engine: a visitor asks a question and gets a written answer built only from your own site content, with citations that link back to the source.

If your site already holds good content but people struggle to find the right page, AI Answers gives them one place to ask in plain language and read a direct answer, instead of scanning a list of search hits. Every answer is grounded strictly in sources the agent retrieves itself, through its own ai_search:rag_search tool or a listing tool you allow, and never from the model's general knowledge, so the response stays true to your site and always shows where each fact came from. It is designed around a "Search Anywhere, Answer Here" pattern, so a search box in your header, a prompt on a landing page and a field in the sidebar can all feed the same shared answer region.

Features

The basic functionality is a question in, a grounded answer out. A Question block posts a visitor's question to the AI Answers API, which runs a retrieval through the agent's tool and composes a written answer from what it finds. An Answer block renders that response together with its cited sources.

• Search Anywhere, Answer Here. Place any number of Question blocks across the site and point them at a single shared Answer block, placed through Block Layout or Drupal Canvas. A Question block can also send visitors to the answer page from any other page, and can show clickable suggested questions.
• Grounded answers only. Responses are built strictly from sources the agent retrieves through its own ai_search:rag_search tool, restricted to a single forced, hidden index so the model can't be prompted into querying elsewhere. A brand-new question always runs a fresh retrieval first; a follow-up question lets the agent decide whether to search again or reuse the previous sources. A turn that ends with no usable sources returns the configured no-answer message. That's guaranteed on the standard JSON response; the live-streaming response relies on the built-in citation instructions to do the same.
• Complete lists. Other agent tools can act as sources, so questions that ask to list or count things ("Which products do you sell?", "How many events are coming up?") get every match from a listing tool, such as Tool Belt's entity list, not just the few most similar search results.
• Citations as rendered entities. Answers carry inline [n] citations that resolve to the underlying content, rendered through a view mode you choose and in the visitor's language when a translation exists. Each source can also link back to a URL of your choice instead of its own page, so citations point at the original when the content was imported or crawled from elsewhere.
• Separate Sources block. An optional block renders the reference list on its own, elsewhere on the page, instead of only inline with the answer.
• Reading content, not chat. Answers are presented as clean, readable output rather than a conversational bubble interface.
• Streaming or single response. The question endpoint can stream the answer as it's written, rendering Markdown to sanitized HTML as it arrives, or return one complete answer.
• Follow-up and feedback. Optional follow-up questions, collapsible conversation turns, and thumbs up/down feedback, saved when AI Logging is enabled and sent to Langfuse as a trace score when Langfuse is enabled.
• Cacheable blocks. The blocks are lightweight, static shells that Drupal can cache. The answer itself arrives over the API via vanilla JavaScript.

Use it when your visitors have questions your content already answers but cannot easily locate: product and service documentation, support and knowledge bases, policy or HR portals, intranets, or any content-rich site where a direct answer beats a list of links.

Post-Installation

The full walkthrough is at Getting started in the full documentation: setting up a vector-database-backed Search API index, configuring an AI Agent's RAG/Vector Search tool with a forced, hidden index, enabling AI Answers for that agent, placing the blocks, and granting the module's permissions to the right roles. It also covers adding a listing tool for list and count questions.

Documentation

Full documentation is published at https://project.pages.drupalcode.org/ai_answers/, including the getting-started walkthrough above, an architecture overview of the request/answer pipeline, and a list of known gaps.

Additional Requirements

Beyond Drupal core, this module requires:

• AI (ai) 1.4 or later
• AI Agents (ai_agents) 1.3 or later
• AI Search (ai_search) 1.3.0-alpha5 or later
• A vector database provider for AI Search, e.g. AI VDB Provider Postgres, and an AI provider that offers embeddings

Recommended modules / recipes / libraries

• AI Recipe: Answers: a recipe for easy setup.
• AI Logging: recommended. Without it, feedback has nowhere to persist.
• Langfuse: optional. Adds tracing for observability, and receives thumbs up/down feedback as a score on the answer's trace (Langfuse 1.0.2 or later). See Optional: Langfuse observability for setup.
• Tool Belt with Tool: optional. Provides the entity listing tool used for list and count questions; see List and count questions.
• crwlr: optional. Sources don't have to be content authored on this site; crwlr crawls an external site into ordinary Drupal nodes on a content type you choose, which can then be indexed and cited by AI Answers like any other content.