Source-linked AI summary
Polaris : Multi Agentic System for Conversational Enterprise Analytics
Varuni H K, Soham Sarkar, Jay Kumar, Goutham Krishnan, Tanvi Johari, Avinash Bharadwaj, Santosh Hegde
TL;DR
Enterprise analytics remains difficult because extracting actionable insights from heterogeneous data requires specialized query, schema, and interpretation expertise, while adaptive coordination for conversational workflows remains underexplored. Polaris addresses this gap with supervisor-led Dynamic Task Coordination and specialized reason-first agents, achieving semantic similarity of 0.85, context precision of 0.99, and answer relevancy of 0.90 across 40 evaluation instances.
Problem
Enterprise analytics remains difficult because heterogeneous data and required query, schema, and interpretation expertise limit accessible insight extraction, while adaptive coordination remains underexplored.
Method
Polaris uses supervisor-led Dynamic Task Coordination to adaptively match specialized agents for querying, visualization, and reasoning under workflow constraints.
Results
Semantic similarity was 0.85, context precision was 0.99, and answer relevancy was 0.90 across 40 evaluation instances.
Takeaways & Limitations
The results suggest principled orchestration is a promising path toward reliable, auditable, and efficient enterprise analytics at scale.
Takeaways & Limitations
Adaptive orchestration with learned utility models and long-term agent memory remains future work.
Abstract
from arXiv · showhide
In today's fast-paced environment, the ability to swiftly access, understand, and act on data is no longer optional; it is essential. Yet most organizations remain data-rich but insight-poor, constrained by the complexity of querying, interpreting, and explaining enterprise-scale information. We present Polaris, a supervisor-led multi-agent framework for conversational enterprise analytics that bridges this gap. Polaris introduces Dynamic Task Coordination (DTC), a decision-theoretic orchestration layer that models agent-task assignment as adaptive bipartite matching, enabling real-time coordination, recovery, and optimization across specialized agents for querying, visualization, and reasoning. By coupling DTC with reason-first, ReAct-style agents, Polaris transforms natural-language queries into coherent analytical workflows that not only retrieve and visualize data but also explain the underlying "why." Evaluation on structured enterprise datasets demonstrates high semantic fidelity and answer relevancy, underscoring the potential of multi-agent orchestration to deliver trustworthy, end-to-end business intelligence at scale.
Introduction
Polaris addresses the barriers posed by complex enterprise analytics and the limitations of single-agent systems through a supervisor-led multi-agent framework. Its Dynamic Task Coordination layer adaptively assigns tasks while integrating data analysis, visualization, and causal reasoning in one conversational interface.
- Motivation: Enterprise analytics is hindered by heterogeneous data complexity and reliance on expertise in query languages, schema understanding, and statistical interpretation.These barriers limit the extraction of actionable insights from large-scale enterprise data repositories.
- Problem: Single-agent LLM systems struggle with compositional reasoning, multi-step coordination, and session-level coherence across enterprise analytics tasks.These tasks include query generation, statistical analysis, visualization, and narrative explanation.
- Problem: Multi-agent enterprise analytics remains underexplored in adaptive task allocation and intelligent error recovery for conversational workflows.Prior multi-agent systems can outperform single agents in tool use and long-horizon planning, but enterprise analytics presents additional coordination needs.
- Contribution: Polaris introduces DTC, a decision-theoretic online orchestration layer that models coordination as adaptive bipartite matching over agents and tasks.Each agent has a capability vector, each task has a descriptor, and edge utilities combine capability fit, empirical reliability, and cost.
- Contribution: DTC selects feasible agent–task matchings that maximize total utility while respecting precedence, capacity, and tool/data-availability constraints.The formulation supports principled allocations by maximizing capability fit and empirical reliability while penalizing cost.
- Contribution: Polaris unifies structured data analysis, dynamic visualization generation, and causal reasoning within a conversational interface.This multi-modal integration targets the gap between data presentation and explanatory understanding in business intelligence systems.
Proposed Work
Polaris is a supervisor-led multi-agent system coordinated by Dynamic Task Coordination (DTC), which adaptively routes specialized agents as goals, latencies, constraints, and data availability change. Its reason-first agents support coherent workflows spanning query generation, visualization, reasoning, and report compilation.
- Dynamic Task Coordination: DTC maintains persistent state across sessions and enforces semantic alignment among visualizations, query results, and narratives.Coordination adapts as user goals, tool latencies, and data availability evolve.
- Agentic Reasoning: Reason-first agents follow ReAct thought–action–observation loops, selecting tools after contextual reasoning and adapting multi-step workflows to intermediate outcomes and system constraints.This approach interleaves chain-of-thought reasoning with concrete actions.
- Dynamic Task Coordination: The supervisor parses intent, maintains context, decomposes ambiguous requests, and routes tasks through capability-aware matching under precedence and capacity constraints.Real-time diagnostics update q(a, t) and constraints, enabling re-matching, rollback, or alternative tools.
- Specialized Experts: The Query Expert converts ambiguous natural language into contextually faithful SQL++ queries using schema inference, semantic annotation, canonicalization, and reasoning-driven aggregation.These mechanisms improve resilience to imperfections in real-world enterprise data.
- Specialized Experts: Specialized experts transform query outputs into semantically appropriate visualizations, generate causally informed explanations, and compile insights into structured reports with methodologies and metadata.The Charting Expert combines rule-based heuristics with ReAct, while the Reasoning Expert uses domain-specific database knowledge.
Results
Polaris was evaluated on a 40-instance golden dataset derived from Airbnb listings in New York. It achieved high average semantic similarity, context precision, and answer relevancy, with threshold results showing consistently precise retrieval and mostly relevant answers.
- Evaluation Setup: Polaris was evaluated using a synthetic 40-query–answer golden dataset built from structured Airbnb rental metadata across New York.Each instance included a natural-language query, reference answer, ground-truth context, and retrieved context.
- Aggregate Results: 0.85 semantic similarity, 0.99 context precision, and 0.90 answer relevancy were achieved across all 40 evaluation instances.These are the aggregated average scores reported for the three evaluation metrics.
- Robustness: 100% of samples exceeded the thresholds for semantic similarity and context precision, while 92.5% exceeded the answer relevancy threshold.Threshold-based evaluation was used to further analyze robustness.
- Interpretation: NL2SQL++ reliably captured user intent and generated precise SQL, reflected in perfect threshold performance for semantic similarity and context precision.The paper attributes the small number of lower answer-relevancy cases to stylistic or explanatory deviations from generative language models despite correct grounding.
Future Work and Conclusion
Polaris is presented as a supervisor-based multi-agent framework for conversational enterprise analytics centered on Dynamic Task Coordination. Future work targets personalization, data-catalog integration, and more editable, exploratory workflows.
- Future Work: Future work includes learned utility models and long-term agent memory to personalize U(a, t | s) for each organization and user.
- Future Work: Future work also includes integrating a global data catalog to maintain consistent, meaningful annotations across varied columns.The goal is to maintain the quality of insights generated by AI agents.
- Future Work: Additional workflow directions include editable plans, dry-run sandboxes, and assisted what-if analyses.
- Conclusion: Polaris formalizes orchestration as adaptive bipartite matching and couples it with reason-first agents for querying, visualization, and explanation.The framework executes coherent multi-step workflows with strong grounding.
- Conclusion: The evaluation shows high semantic similarity and context precision, indicating faithful retrieval and reliable answer generation.