In the present days, multiple devices generate large amount of data such as communication, transaction, healthcare and IoT etc. continuously over time. Topic modelling is a suitable technique for analysing this amount of large text datasets without too much overhead. This motivates the implementation of a topic modelling method that can capture the optimize topics in an adaptive manner. Therefore, the paper proposes CA-BERTopic, a collaborative and adaptive framework for transformer-based topic modelling built upon the BERTopic architecture. The framework enables multiple distributed data sources to collaboratively contribute to topic discovery without sharing raw data, ensuring privacy and scalability. Unlike conventional centralized models, the proposed approach adapts iteratively and refines topic models as new data becomes available. This adaptivity allows the model to capture the natural evolution of topics across time while maintaining coherence and interpretability. The framework has been evaluated on two datasets, a Long-Form Question Answering corpus involving doctor–patient interactions and a Tweets dataset. Experimental results demonstrate that the proposed method achieves comparable performance to centralized BERTopic and performs better than existing methods of LDA, Guided LDA, Ensemble LDA and LDA based on BERT embeddings.