A Multi-Agent Framework for Complex Product Retrieval over Enterprise APIs

Sammanfattning

Axis Product Information API (PIA) is an internal REST interface exposing product hierarchy, specifications, compatibility relations, and lifecycle links. This thesis designs, implements, and evaluates a multi-agent LLM system that lets users query this structured product data in natural language. Three experiments contrast (RQ1) orchestrator–worker versus single-agent decomposition, (RQ2) on-demand hierarchy traversal versus LLM-driven traversal, and (RQ3) two-tier reflection versus no reflection. The traversal layer removes traversal-side LLM calls and reaches 100% validity on deep traversals where the historical LLM-recursive baseline reaches 0%, while remaining statistically indistinguishable overall from a tuned one-shot LLM baseline. Reflection-mediated recovery, with predefined clarification supplements, improves validity by 20 percentage points at a +30% token cost. The orchestrator–worker design does not outperform a single agent on pooled validity but uses fewer tokens and provides the substrate for both. The findings support bounded, graph-inspired orchestration with explicit recovery gates for enterprise API agents.