The problem
Two public systems had to answer a person, not an analyst. One helped voters find a polling station and a voting date across on the order of hundreds of millions of records. The other searched health-insurance beneficiaries, with grouping and autosuggest so a near-match still led somewhere.
The approach
Both sat on Apache Solr. The useful design was the lookup a person could finish, including bilingual web search on a separate state site, not a cluster diagram.
The conclusion
At that size the interface is the product. If the person cannot tell they found the right row, the index does not matter.