How we built grounded AI assistants for HOMELand Lebanon and CASPR in Dublin, and what we learned about retrieval, trust and handover to humans.
Most AI chatbot demos answer from general knowledge. Real organisations need something different: answers grounded in their own content, with a clear path to a human when the bot shouldn't answer.
This post covers two projects where we built exactly that. One is a cultural magazine archive in Lebanon, the other a community education charity in Dublin. The content, the users and the risks were very different, and that is where most of the lessons came from.
Project 1: HOMELand Lebanon, a living archive made conversational
The problem
HOMELand has been publishing since 2014 (formerly HOME Magazine) as a cultural intelligence platform about Lebanon and its diaspora. The archive is broad:
- Seven content verticals: THEdiaspora, THEland, THEpeople, THEculture, THEdiplomacy, THEfuture and THEgood
- Ten magazine issues
- Country pages, interviews, writer profiles and videos
Readers usually arrive looking for one specific thing: a person, a village, a family, a story from a particular country. When keyword search can't find it, they leave.
What we built
Jeddo, an AI guide grounded in the archive. It answers in plain language, links to the article or issue behind each answer, and opens as a chat widget on any page of the site.
Where it stands
Deep archive pages are now as reachable as the homepage, through a single conversation.
Project 2: CASPR, answers for parents, volunteers and funders
The problem
CASPR, the Community After Schools Project, has served Dublin's North East Inner City since 1995. Its mission is to counter educational disadvantage in the community. In practice that means:
- An after-school programme across three locations, serving 150 primary school children from five local schools
- Community Employment with QQI-certified training in childcare and ECDL for 36 local people
- Work with community and statutory agencies on care plans for children most at risk
- A steady need for volunteers and donations
The same questions keep coming in from parents, volunteers and funders, and staff answer them between sessions.
What we built
A virtual assistant trained on CASPR's services, programmes and published resources. It helps each visitor find the right service in their own words before they pick up the phone.
Where it stands
Repetitive reception questions are handled instantly, and staff time goes back to the children.
What we learned
- Grounding beats a bigger model. Restricting answers to retrieved content and showing sources did more for trust than any prompt tweak.
- Archive users and service users search differently. HOMELand readers search by names and places. CASPR visitors describe problems in everyday words. Retrieval has to handle both styles.
- Design the "I don't know" path first. A good fallback matters more than a clever answer, especially for a nonprofit.
- Keep the AI layer decoupled. Running the chatbot as a separate service and embedding it as a widget made it easy to add to existing WordPress sites.
Want something similar?
Both projects are live: HOMELand Lebanon and CASPR.
At Techowdy we build custom AI chatbots trained on your business documents, FAQs and website content, for the web and WhatsApp. You can read more about our AI chatbot development services.
If you've built a RAG chatbot over an existing content library, what was the hardest part? I'd like to hear in the comments.


Top comments (0)