BizSage: A Self-Evolving Multi-Agent Framework for Business Research with Efficient Knowledge Retrieval

BizSage is a self-evolving multi-agent framework for economics and business research that utilizes a Lateral Knowledge Graph (LKG) to address the granularity mismatch of paper-level retrieval. By merging section-level knowledge graphs from over 130,000 papers, it enables fine-grained retrieval using Personalized PageRank (PPR) and cosine semantic matching to surface specific, structurally important evidence.

The system employs seven specialized agents collaborating under a Meta-Review self-evolution mechanism that distills evaluation feedback into reusable strategies. This approach allows BizSage to achieve leading scores on research benchmarks, maintain zero hallucinated citations, and adapt to domain-specific quality standards through iterative refinement.

BizSage is presented as a self-evolving multi-agent framework for business research that pairs an agent-based reasoning layer with a pre-built, section-level scholarly knowledge graph. Its central aim is to make literature-grounded business research more efficient and more precise than approaches that retrieve from flat document corpora or from knowledge graphs organized primarily around papers or high-level concepts. By constructing the knowledge graph offline at the granularity of paper sections, the framework preserves finer-grained scholarly units—such as problem definitions, methods, findings, and contextual arguments—rather than treating an entire paper as a single retrieval target.

The key retrieval insight is that effective business-research support requires both semantic similarity and structural relevance. BizSage combines cosine semantic matching with Personalized PageRank propagation over the section-level graph: cosine matching identifies directly relevant sections, while PPR spreads relevance through the graph to related sections, adjacent arguments, and connected scholarly context. This hybrid design is intended to overcome the limitations of concept-level or paper-level knowledge graphs, which can be too coarse, ambiguous, or disconnected from the specific evidence needed for a research question. The multi-agent component is positioned as the “self-evolving” layer, suggesting that the system can iteratively refine its research planning, retrieval, synthesis, or validation behavior while relying on the offline graph as a stable, reusable knowledge substrate.

The material matters because it offers a scalable architecture for domain-specific research assistants, especially in fields like business research where questions are often multi-hop, conceptually dense, and require grounded synthesis across many scholarly sources. By moving expensive graph construction offline and keeping online retrieval relatively lightweight, the framework targets a practical tradeoff between depth of reasoning and operational efficiency. More broadly, it points toward a useful pattern for advanced RAG systems: pre-structuring knowledge at an appropriate semantic granularity, combining dense matching with graph-based propagation, and using agents not merely as wrappers around retrieval but as adaptive components that improve the research workflow over time.