Answers are grounded in this paper and cite their sources.
Generative artificial intelligence (GAI) systems increasingly support extended interactions involving programming, scientific research, mathematical computation, data analysis, software development, and other complex tasks. Throughout these interactions, users and AI systems continuously generate messages containing instructions, questions, explanations, source code, computational parameters, experimental configurations, intermediate results, uploaded files, and other potentially reusable information. However, retrieving relevant information from extensive interaction histories remains a significant challenge, particularly when the information required for a current computational task is distributed across different conversations, dates, message senders, and attachment types. Conventional conversation search primarily emphasizes textual matching or conversation-level retrieval, while existing conversational retrieval and agent-memory approaches do not necessarily provide a unified, user-configurable mechanism that connects multidimensional message filtering directly to computational task preparation and execution. This paper proposes Task-Oriented Bidirectional Message Search and Filtering for Generative Artificial Intelligence (TBMSF-GAI), a framework that treats both user-generated and AI-generated messages as independently searchable information units within a unified interaction archive. The proposed framework supports multidimensional filtering according to message sender, date and time, keywords, and file types, together with optional semantic retrieval, cross-conversation searching, compound filtering, attachment identification, contextual reconstruction, and task-oriented information assembly. Unlike approaches that search only user inputs or retrieve entire conversations without sufficiently distinguishing individual message attributes, TBMSF-GAI establishes a bidirectional retrieval model in which messages from both interaction participants can be selected, examined, ranked, and assembled according to the requirements of a computational task. Its central contribution is the conceptual integration of message-level retrieval, structured filtering, computational relevance assessment, and context construction into a unified task-oriented process. To the best of our knowledge, this paper advances a first proposed unified framework specifically defining this combination as a distinct computational capability for generative AI systems. This novelty claim concerns the proposed integration and task-oriented architecture, rather than the individual search and filtering techniques, which have established precedents. The paper presents the conceptual architecture, formal model, operational workflow, implementation considerations, evaluation methodology, and research implications of the framework.
No takes yet. Share an insight, caveat, or question.
Jincheng Zhang (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: