AI Answers¶

AI Answers turns an existing AI Agent into a RAG-powered answer engine. An "answer" is a written response generated by the agent, grounded strictly in sources the agent retrieves itself through its own ai_search:rag_search tool, with inline [n] citations linking to those sources rendered as Drupal entities.

Key features¶

Live, streamed answers: text appears as the agent writes it, instead of waiting for the full answer.
Settings per agent: each AI Agent controls whether it's allowed to answer, which model it uses, how sources are shown, and what to say when it can't find one.
Follow-up questions: visitors can keep asking in the same conversation, with a limit on how much of the earlier chat gets reused.
Thumbs up / down feedback: visitors can rate an answer, and that feedback can optionally feed into Langfuse for observability.
Example questions and flexible placement: the question box can show clickable example questions, and can still reach its answer block even if the two aren't on the same page.
Optional debugging details: site builders can grant a permission that shows extra tracing information alongside answers, for troubleshooting.

How it works¶

flowchart TD
    subgraph FE1["Browser: Question and Answer blocks"]
        A["1. Visitor types a question"]
        B["2. The Answer block sends it to the server"]
        A --> B
    end

    subgraph BE["Server: building the answer"]
        C["3. AI Answers picks up the request"]
        D["4. Checks the AI Agent is<br/>allowed to answer and has a<br/>way to search for sources"]
        E["5. Searches for sources and<br/>keeps only the good matches"]
        F["6. Asks the AI Agent to write<br/>an answer using those sources,<br/>citing them as [1] [2] ..."]
        G["7. Streams the answer back as<br/>it's written, finalizing the<br/>citations once it's done"]
        C --> D --> E --> F --> G
    end

    subgraph FE2["Browser: showing the reply"]
        H["8. The answer appears live,<br/>with sources listed and<br/>citation markers linked to them"]
        I["9. Visitor can rate the answer<br/>(optional thumbs up/down)"]
        H --> I
    end

    B --> C
    G --> H

    J(["↻ A follow-up question reuses<br/>the same conversation"])
    I -.-> J
    J -.-> B

This is the simplified version. See Architecture for the full pipeline, including the gates, the event-subscriber bridge, and the citation-finalization step referenced by step 7 above.

Sections¶

Getting started: cross-module requirements, enabling the module, configuring an agent for answers, placing blocks, and debugging a run via ai_logging.
Architecture: the request flow, the answer pipeline, the event-subscriber bridge into the agent loop, source processing, and the front-end contract.
Known gaps: limitations that are not yet fixed, as distinct from the accepted architectural trade-offs recorded in the decision log.
Decision log: the canonical record of why the architecture is shaped the way it is. When an architectural choice changes, update the decision log in the same commit. It is the only place this rationale is expected to survive.