AI Agent using Natural Language Processing (NLP) gives users the ability to make complex data queries using natural language, bypassing time-consuming manual data transformation, processing and analysis.
If you ask an AI agent, out of the box, to perform some data analyses or inferences, it’ll go off and do its thing, with very little visibility to its logic and reasoning. Hallucination occurs when the AI generates responses that appear coherent but are factually inaccurate or logically inconsistent. An AI tasked with summarizing project budget, for example, might infer nonexistent data patterns or fabricate correlations if it encounters incomplete or ambiguous input data. It might produce generic advice with little to no practical value based on presuppositions that are irrelevant.
Multi-Layered Agent Architecture
To mitigate this, we adopt a multi-layered agent architecture with stateful reasoning, allowing discrete agents to verify intermediate steps and maintain alignment with the data’s logical structure throughout the analysis process. This multi-layered agent consists of three primary components:
- The Assistant Agent: The Assistant Agent is the interface between the user and the backend data system. Using NLP, it processes natural language queries to extract intent, key parameters, and context. It creates a structured blueprint for the downstream State Manager and Query Agents to analyze relevant data, ensuring complex queries are systematically addressed. For instance, a query like “What’s the status of the current project budget?” is parsed into its core components: “status,” “project budget,” and “current.” The Assistant Agent does not conduct analyses on the data directly (unless the query is trivial) but provide structured instructions for specialised agents to carry out analyses.
- State Manager: The State Manager is responsible for managing a query parsed from the Assistant Agent and decides how the query is to be processed by parsing and distributing instructions from the Assistant Agent to specialised Query Agents synchronously or asynchronously. It monitors the query state and determine when the query has been met or if there is insufficient information in its knowledgebase.
- The Query Agents: A Query Agent takes inputs, context and prompts from the State Manager and performs the necessary processing and analysis. Each agent specializes in a specific type of query, such as compiling schedules, assessing risks, or evaluating financials. Once a query has been completed, the agent returns its response back to the state manager in a structured format for further processing. During each query iteration, the agents can output its reasoning and confidence level to the State Manager and for user evaluation and troubleshooting as necessary.
A Query Agent is configured to adopt Multi-Step Reasoning, Multi-shot Prompting, Chain of Thought, Dynamic Interactive Reasoning and other advanced techniques. Additionally, a Query Agent can read external datasets, context and call upon predefined data workflow and logic encoding to perform sematic and inference analysis. Structured inputs and outputs are passed between Query Agents via the State Manager. This layered architecture allows the State Manager to segment a query into manageable tasks to follow logical steps during the process.
This multi-layered approach empowers the AI system to adapt to diverse query types while maintaining systematic, logical, and explainable operations. It allows the AI to evaluate user queries intelligently, decide on the best analytical approach, and execute tasks with precision, ensuring that the results meet user expectations.
