As　large　language　models　advance,　human-machine　interaction　has　become　increasingly　common,　yet　challenges　remain　regarding　their　adaptability　to　specific　populations　and　scenarios.　Focusing　on　the　emotional　needs　of　the　elderly,　this　study　develops　a　customized　retrieval-based　dialogue　generation　system　for　companion　robots.　Through　dialogue　testing　and　data　cleansing,　a　lightweight　system　is　built　and　evaluated　using　both　objective　and　subjective　measures.　Results　show　improved　response　usability　and　recommendation　rates.　Compared　to　traditional　neural　network　training,　this　method　offers　simpler　implementation,　faster　deployment,　and　effective　performance,　making　it　a　practical　complement　to　large　language　models.　©　The　Institution　of　Engineering　&　Technology　2025.