ResearchFlow: A Multi-Agent AI-Driven Platform for Systematic Literature Review Generation

ResearchFlow is a commercial-grade multi-agent AI platform that assists researchers in automatically generating scoping reviews and systematic reviews via human-machine collaborative workflows. The platform's core philosophy is "researcher-led tooling", which retains human confirmation at key decision points to balance AI efficiency with human autonomy. It integrates multi-agent architecture, Retrieval-Augmented Generation (RAG), and other technologies to address the pain points of traditional literature reviews, such as time-consuming processes and susceptibility to bias.

Traditional manual scoping/systematic reviews face issues like low efficiency, easy omission of literature, or introduction of selection bias. The ResearchFlow project emerged to automate this process through a multi-agent AI system and human-machine collaborative workflow, while maintaining the decision-making autonomy of human researchers.

Positioned as a commercial-grade tool, the platform's core features include conversational planning, automatic search string generation, transparent operations, RAG intelligent assistant, and cloud-native architecture. The architecture centers around three agent clusters: the Research Cluster (literature scouts, data extractors, etc.) handles literature discovery and analysis; the Writing Cluster (article writers, citation managers, etc.) converts results into academic text; the Quality Cluster (evaluators, fact-checkers, etc.) ensures content quality.

The platform adopts a Human-in-the-Loop design, setting human confirmation points at key stages to retain the dominance of human judgment; users can interact via a conversational interface and review intermediate results. It integrates a hybrid RAG search system that combines semantic search and BM25 keyword search, optimizes results through query decomposers and re-rankers, and uses ChromaDB to store literature for efficient similarity retrieval.

The platform iteratively improves through a saturation cycle mechanism: writing, evaluation, fact-checking, and consistency checking form a closed loop. A section is only completed when its quality score reaches the threshold; otherwise, it is revised based on feedback. Multi-dimensional quality assessment covers dimensions such as accuracy, completeness, consistency, academic style, and citation norms.

It mainly targets academic researchers, research institutions, and corporate R&D departments. For teams that need to write reviews regularly, it can significantly shorten the cycle from retrieval to finalization; the human-machine collaborative design is suitable for academic scenarios with strict quality requirements, and the cloud-native architecture supports on-demand scaling to adapt to projects of different sizes.

Currently, it mainly supports scoping reviews; in the future, it can be extended to more academic writing tasks such as research proposals and technical reports. Direct integration with specific database APIs still needs improvement; potential enhancement directions include real-time collaboration functions and richer visualization options.

ResearchFlow combines multi-agent systems, RAG, and human-machine collaboration to strike a balance between automation and human control, providing an efficiency tool for academic research and promoting a new research paradigm: humans focus on high-level judgment and innovation, while AI handles tedious information processing. With the development of large language models, such intelligent assistants will play a more important role in academia.