To address challenges in the power sector, such as heterogeneous knowledge sources, complex domain-specific semantic relationships, the limitations of traditional retrieval methods in handling multi-turn interactions, and the tendency of large language models (LLMs) to generate factual hallucinations—this paper proposes an interactive intelligent question-answering framework for power-domain knowledge bases. This framework integrates multi-agent knowledge construction, context-aware processing, and adaptive retrieval-augmented generation (RAG). First, an ontology-constrained, multi-agent collaborative knowledge extraction mechanism is designed; through entity localization, attribute fusion, and evidence tracing, a hybrid vector-graph knowledge base is constructed using FAISS and Neo4j. Second, the framework incorporates structured dialogue memory, context-aware query reformulation, and a two-tier keyword enhancement strategy to facilitate entity coreference resolution, semantic completion, and the propagation of domain-specific constraints across multi-turn interactions. Building on this, a multi-path retrieval mechanism is developed, incorporating semantic coverage, terminology anchoring, and document structure priors, which utilizes a lightweight language model to dynamically adjust retrieval weights and optimizes evidence ranking through cross-path consensus feedback. Finally, by integrating evidence-constrained generation, factual consistency verification, and uncertainty assessment, the framework enhances the traceability and reliability of the generated answers. The proposed framework is evaluated through the design of a multi-type power-domain QA benchmark, alongside comparative experiments, ablation studies, and robustness tests, assessing performance across dimensions such as retrieval quality, answer accuracy, multi-turn consistency, and computational efficiency.