Project Overview

This page provides a high-level overview of the project, including its goals, the technologies used, the internal architecture, and a representative use-case scenario.

Project Description

This project investigates the integration of Large Language Models (LLMs) into an agent-based system capable of performing complex reasoning across various domains. The architecture is modular, explainable, and designed to combine autonomous language understanding with external tool access and iteration over structured workflows.

### Key components and technologies: - LangGraph and LangChain for graph-based agent orchestration. - Vertex AI (Gemini) as the main inference engine. - Open-Meteo API for historical and forecasted weather data. - Elsevier API and Crossref API for retrieving scientific literature. - InfluxDB for querying time-series sensor data from laboratories. - Custom agents for handling tasks such as weather retrieval, database access, scientific search, and numerical calculations. - Supervisor agent for iteratively coordinating agent calls and ensuring task completeness. - Sphinx and ReadTheDocs for documentation generation and deployment.

Data Flow Architecture

The system accepts a complex user query in natural language. A Supervisor component processes the query and dynamically routes it through a graph of specialized agents. Each agent performs a well-defined sub-task, such as retrieving sensor data, calculating averages, or searching for scientific models.

The execution graph enables iterative and conditional decision-making: if more information is required, the supervisor can invoke additional agents or reprocess previous results. Once all necessary steps have been completed, the system synthesizes the outputs and delivers a final, explainable response.

The following diagram illustrates the overall architecture and agent interaction:

Example Scenario

> “Estimate the heat generation produced by users in Laboratory A from the 5th to 9th of December 2024. Use the temperature difference between indoor sensors and the outdoor temperature in Madrid. Since no occupancy data is available, assume working hours between 08:00 and 18:00 on weekdays. Then search for scientific models of heat generation. Extract the most recent model’s formulation and compare it with your method.”

This prompt triggers the following system behavior: - The Supervisor identifies the need to retrieve both indoor and outdoor temperature data. - The DBAgent collects indoor sensor data from InfluxDB within the specified date and time range. - The WeatherAgent fetches outdoor temperature data from Open-Meteo. - The CalculatorAgent computes the differential heat profile and estimates human heat generation using known physical assumptions. - The ElsevierAgent and CrossrefAgent search for recent scientific models related to heat generation in indoor environments. - The system compares the computed values with the extracted scientific formulation and generates a reasoned response with quantitative and qualitative analysis.

This scenario demonstrates the system’s capacity to chain multiple agents in a supervised loop, integrate structured data with domain knowledge, and generate results with traceable and explainable logic.

Example Scenario Execution

Prompt: “Estimate the heat generation produced by users in Laboratory A from the 5th to 9th of December 2024. Use the temperature difference between indoor sensors and the outdoor temperature in Madrid. Since no occupancy data is available, assume working hours between 08:00 and 18:00 on weekdays. Then search for scientific models of heat generation. Extract the most recent model’s formulation and compare it with your method.”

Execution Summary: - The Supervisor decomposed the task into a multi-step reasoning plan, aiming to retrieve temperature data, compute temperature differences, and retrieve scientific models for comparison. - The DBAgent successfully retrieved indoor temperature data from InfluxDB for the specified period and time range. - The system attempted multiple queries for outdoor temperature data using both the WeatherAgent and InfluxDB, but encountered limitations due to missing or misconfigured location metadata. - As fallback, the system proceeded with an estimation based on typical human metabolic rates, introducing a range of assumptions to cover different occupancy and activity levels. - The CrossrefAgent and ElsevierAgent were unable to retrieve a valid scientific model within this execution. - The VerifierAgent evaluated the final output and assigned a satisfactory verdict with a confidence score of 0.6, acknowledging that the fallback reasoning was reasonable given the data constraints.

Final System Output:

Due to the inability to retrieve specific temperature and occupancy data, the estimated heat generation in Laboratory A during working hours (8:00 - 18:00) from December 5th to 9th, 2024, is estimated to be between 140 W and 1500 W. This range is based on the following assumptions:

- Occupancy: Between 2 and 10 people.
- Metabolic Rate: Between 70 W/person (resting) and 150 W/person (active work).

It is important to note that this is a rough estimate. The actual heat generation will depend on the actual occupancy and activity levels in the laboratory. Access to sensor data and scientific models would significantly improve the accuracy of this estimation.

The VerifierAgent returned a confidence score of 0.6 and a "satisfactory" verdict.

Observations: This scenario illustrates the system’s ability to handle degraded conditions gracefully: - It correctly followed the reasoning plan and retrieved available data. - It transparently reported missing data and adapted its reasoning accordingly. - It introduced reasonable fallback assumptions with explicit disclosure. - The final output was validated and explained, maintaining system transparency.

Such robustness in handling complex, multi-step prompts under partial data availability highlights the explainability and adaptability of the agent-based architecture implemented in this project.