Ask Susinsight: a retrieval grounded research assistant for African sustainability reporting

Susinsight publishes research and reporting on sustainability and development across Africa, but readers had no fast way to pull answers out of that archive. Search returns a list of stories, not an answer, and a generic chatbot would happily invent facts the publication never reported. The assistant had to feel useful without pretending to know more than Susinsight had actually covered, and every answer needed to trace back to a real published story so a reader could check it. The hard part was trust: an assistant that fabricates a source or overstates thin evidence damages the publication's credibility more than having no assistant at all.

I led the design and implementation of Ask Susinsight across two surfaces: a dedicated /ask page and a floating modal available site-wide. A reader asks a question, the system retrieves relevant article passages, and the answer cites and links back to the original stories. The backend uses retrieval-augmented generation. Articles are chunked, embedded, and searched by similarity. Matched passages are passed into the language model with strict instructions to answer only from the retrieved material, and to say so plainly when the archive does not cover the question. Most of the value came from the edge cases. I tightened source handling so article titles could never be invented, added query expansion so broad questions still matched relevant passages, and rewrote the prompt rules for moments when the assistant tried to stretch thin evidence too far. The assistant also supports multiple languages, so readers across the continent can ask in their own words.

Ask Susinsight turns a growing archive into something readers can question directly, with answers that link straight back to the reporting behind them. Because the model answers only from retrieved passages and cannot invent source titles, a reader can verify every claim against a real story. The clearest takeaway was that retrieval quality matters more than model size: better metadata, clearer prompts, and honest fallback behavior when evidence was thin made the assistant more trustworthy than swapping in a larger model would have. The result extends the publication's credibility instead of putting it at risk.