A novel incremental graph construction technique has been introduced to significantly bolster the robustness of spectral clustering when applied to text embeddings. Conventional k-Nearest Neighbor (k-NN) graphs, which are foundational for this analytical approach, frequently present a critical vulnerability: they often contain numerous disconnected components within realistic text datasets, especially at practical sparsity levels with small k values. This inherent fragility renders spectral clustering degenerate and highly sensitive to hyperparameter tuning, compromising the reliability of its output. The newly proposed method directly confronts this challenge by employing an incremental construction strategy for k-NN graphs that explicitly preserves graph connectivity1. By fostering a more integrated and robust graph topology, this approach effectively mitigates the issues of instability and hyperparameter dependency that plague standard k-NN graph generation. For practitioners dealing with the complexities of textual data, this advancement offers a more reliable and less arduous pathway to effective text analysis and grouping, leading to more dependable insights from large language datasets.