When the question uses different words
A visitor may describe a question differently from the source material. Search needs to find relevant content and give the visitor a way to inspect it. The sources need to support any generated answer.
What I delivered
I delivered the House of Destiny WordPress AI search beta, covering retrieval design and application engineering. It combines OpenAI embeddings and Supabase vector retrieval with keyword matching. The application adds summaries and filters, plus caching and analytics.
How retrieval works
Vector retrieval finds material with a similar meaning. Keyword matching finds words that appear in the source. Combining them supports questions phrased differently from the material while still handling searches for specific terms.
Summaries use retrieved content and link back to its sources. Filters narrow the search context, with caching and analytics supporting the application around retrieval.
What still needs checking
The next evaluation needs representative questions and exact-term searches, including cases where sources are missing. It also needs to check whether the retrieved material supports each summary. Those checks are requirements for evaluation. They are not completed benchmark results.
Other search projects
JumpCloud AI search is a separate working prototype and business case. Relatable is a separate WordPress semantic-search plugin using local vectors and adaptive hybrid ranking. Each has its own implementation and release status.
Discuss the work
Tell me what your team needs to build, or where the work is getting stuck.