Drupal AI on Halcyon

Halcyon Stellar Works is a made-up shipyard, but everything around it is real Drupal 11. The same ~420 pages are searched three different ways, and an AI assistant, editor tools, automation and translation run on the same content. Every AI request goes to amazee.ai: chat models, embeddings and the vector database, all through the Drupal AI module and its amazee.ai provider.

Four ways to search the same content

Search has moved from matching words to matching meaning, and from returning links to answering questions. Each approach still has its place. Run one query through the first three at once on the compare search page.

Keyword search with Solr

Search API + Apache Solr · lexical

The traditional approach. Every page is split into words and stored in an inverted index on a Solr server. A query returns pages that contain the words you typed, ranked by how often and where they appear (title matches count more).

Advantages

  • Exact and predictable: names, part numbers and codes are found every time.
  • Fast and cheap at scale; no per-query AI cost.
  • Mature: facets, highlighting, spellcheck, synonyms, stemming.
  • Easy to explain why a result matched.

Disadvantages

  • Needs the right words: "cheap ship for one pilot" finds nothing if pages say "low-cost single-seat courier".
  • Every word must match by default, so longer natural questions return zero results.
  • Synonyms and tuning are manual work.
  • Another server to run and keep in sync.

Best for: known-item lookup, catalogues, admin search, legal or technical text where exact terms matter.

Try “Atlas Heavy Freighter”

Scolta: browser search with AI on top

Scolta by Tag1 · Pagefind index · amazee.ai

Drupal exports a static, compressed index (Pagefind) that the visitor's browser downloads in small pieces and searches locally in WebAssembly. Scolta re-ranks the results and asks an LLM, through Drupal AI and amazee.ai, to expand the query with related terms, write a short overview with links, and answer follow-up questions.

Advantages

  • Very fast: search runs in the browser, no search server.
  • Natural language works, because AI expansion adds the words you did not type.
  • AI overview and follow-ups turn results into answers, grounded in the hits.
  • Works without AI too: plain keyword search when no provider is set.
  • Facets and AI-chosen filters and sorting ("cheapest", "near Europa").

Disadvantages

  • Still lexical underneath: meaning comes from expansion, not from embeddings.
  • The index is rebuilt by Drupal on content changes (queued, not instant).
  • Everything in the index is public; not for access-controlled content.
  • AI overview adds an LLM call and some latency.

Best for: public sites that want fast, forgiving search and AI answers without running a search cluster.

Try a natural question

Semantic search with a vector database

AI Search · amazee.ai embeddings · pgvector

Each page is split into chunks, and an embedding model turns each chunk into a long list of numbers that captures its meaning. These vectors live in amazee.ai's Postgres/pgvector database. A query is embedded the same way, and the closest chunks by cosine similarity are returned, whatever words they use.

Advantages

  • Matches meaning: "ore", "mining hauler" and "Belt refinery runs" all land together.
  • Robust to phrasing, typos and synonyms without manual tuning.
  • Cross-lingual by nature with multilingual embedding models.
  • The foundation for RAG: the same index feeds the chatbot below.

Disadvantages

  • Each query needs an embeddings call: slower (seconds, not milliseconds) and has a cost.
  • Indexing is slow and paid: every chunk of every page is embedded.
  • Weak on exact identifiers and rare names; scores are hard to explain.
  • Always returns "nearest" results, even when nothing is truly relevant.

Best for: research, support and knowledge bases where people describe what they need rather than name it.

Try a fuzzy question

Answers, not links: RAG chatbot

AI Assistant API · AI Chatbot · retrieval-augmented generation

Ask Halcyon (bottom right of every page) searches the vector database for the chunks most relevant to your question, then gives them to an LLM with instructions to answer only from that context and link the pages it used. It keeps the conversation, so you can ask follow-ups.

Advantages

  • One direct, conversational answer that combines several pages.
  • Grounded in your own content, with links back to the source.
  • Instructions shape tone, scope and what it refuses to answer.

Disadvantages

  • Slowest option: retrieval plus generation, several seconds per answer.
  • LLM cost per message; needs rate limits on a public site.
  • Can still misread or over-summarise; answers must be checked against the links.

Best for: guided discovery, product advice and support questions.

Open “Ask Halcyon” and try: Which ship can a single pilot fly to Europa?

Side by side

Solr keyword Scolta Vector RAG chatbot
Natural-language questionsPoorGood (AI expansion)Very goodVery good
Exact names and codesExcellentExcellentFairFair
Typical response time~10–300 ms~100–600 ms + overview~1–7 s~5–20 s
AI cost per queryNoneOptional (expansion, overview)1 embeddings callEmbeddings + LLM
InfrastructureSolr serverStatic files, no serverVector DB + embeddings APIVector DB + embeddings + LLM
Indexing speed (this site)Seconds~1 second~40 minutesShares the vector index
Explainable rankingYesMostlySimilarity score onlyVia cited links
Facets and filtersYesYes, AI can pick themMetadata filtersNo
OutputRanked linksLinks + AI overviewRanked passagesA written answer with links

In practice, combine them: keyword for precision, vectors for recall (AI Search can also boost a Search API database or Solr index with vector results), and RAG on top when people want an answer.

More Drupal AI on this site

Search is one use of Drupal AI. These run on this site too. Features marked Editor need a login; on a trial or a local copy, use drush uli.

Try it now

Ask Halcyon chatbot

A RAG assistant built with the AI Assistant API and AI Chatbot (Deep Chat) modules, answering from the vector index with links to sources.

Try it: open the chat bubble, bottom right, and ask “Who flies Leviathan Ore Haulers, and how many?”

Try it now

AI search overview and follow-ups

Scolta sends the top results to an LLM through Drupal AI for a short, linked overview, and lets you ask follow-up questions about them.

Try it: search, read the AI overview in the Scolta column, then ask a follow-up.

Open compare search
Editor

AI in the editor (CKEditor)

An AI button in the CKEditor toolbar: continue writing, summarise, change tone (including “Mission briefing”), translate, fix spelling, reformat HTML, or run your own prompt on selected text.

Try it: Content › Add content › Blog post, select a paragraph, click the AI button.

Add a blog post
Editor

AI Automators

Fields that fill themselves. Blog post tags are chosen by an LLM from the body when you save, reusing existing tags where they fit. Automators can also write summaries, extract data, generate images or chain steps.

Try it: save a blog post with an empty Tags field, then look at its tags.

Editor

AI Content Suggestions

On every edit form: suggested titles, a summary, a readability check, tone analysis, content moderation, and taxonomy term suggestions for the text you wrote.

Try it: edit any ship or post and open the AI suggestions panel.

Edit the Atlas Heavy Freighter
Editor

AI Translate

One-click translation of whole pages into German, French and Japanese, as real Drupal translations you can review and edit.

Try it: open a page's Translate tab and choose “AI Translate”.

Translate a ship page
Admin

AI API Explorer

A playground for site builders: try chat, embeddings and other operations against any configured provider and model, and see the code to call them from PHP.

Open the explorers
Admin

AI Logging

Every prompt this site sends is logged with its provider, model and tags, so you can see exactly what goes to the AI and what each feature costs.

View the AI log
Admin

One provider: amazee.ai

Chat, embeddings and the vector database all come from the amazee.ai provider. Features ask for model aliases like chat and embeddings, keys live in the Key module, and switching provider is one setting.

AI default providers

What else Drupal AI can do

Not installed on this demo, but part of the same ecosystem:

  • AI Agents: agents that build content types, fields and views, or edit content, from a chat prompt.
  • Image AI: alt text generation, image classification and image generation in automators.
  • ECA + AI: no-code workflows where an AI step classifies, routes or rewrites content on events.
  • Field widget actions and AI validations: AI buttons next to any field, and validation rules written in plain language.
  • External moderation: screening of user-submitted text before it reaches an LLM.
  • MCP: exposing Drupal content and actions to external AI tools.