A number of interfaces have been proposed in recent years to help users build SPARQL queries, including textual editors with syntax highlighting and error correction, and visual editors that allow for drawing graph patterns using node and edge components. A common feature supported by such systems is autocompletion, which offers users suggestions for terms to insert into a query, potentially restricted by a keyword prefix. However, current systems either return irrelevant terms that will generate empty results, or return relevant terms but may time out while generating suggestions for complex queries. We propose an autocompletion technique based on a graph summary that aims to strike a balance by over-approximating relevant results in an efficient manner.