Abstract
The CALO (Cognitive Assistant that Learns and Organizes) project aims to create an intelligent assistant capable of managing complex tasks by integrating diverse, heterogeneous knowledge sources. This paper presents the CALO Query Manager, a system that utilizes a hybrid reasoning architecture to orchestrate these sources. By treating each source as a specialized reasoner and employing a sophisticated query planner, the system decomposes high-level user queries into executable logical plans. Evaluation demonstrates that the Query Manager handles complex queries with minimal overhead, typically completing planning in under 200ms.
Introduction
Building a truly intelligent assistant requires the ability to aggregate and reason over data from a wide variety of sources, including emails, public schedules, sensor feeds, and live web data. These sources are inherently heterogeneous: they use different schemas, possess varying levels of reasoning capability, and often impose strict binding constraints (e.g., requiring a specific input to provide an output). Traditional systems often treat these as passive databases, which limits their flexibility. The CALO Query Manager addresses this by treating every source as an active reasoner and providing a central management layer to coordinate them through a mediated ontology.
System Architecture
The Query Manager is built upon the Java Theorem Prover (JTP) architecture and incorporates several novel components to manage the complexity of multi-source reasoning.
Key components include:
- Mediated Schema: A shared CALO Ontology ensures that all disparate sources communicate using a common vocabulary.
- Rule Expansion Reasoner: This component applies Horn clauses to expand initial queries. For instance, a query about a person's presence might be expanded into subgoals involving camera feed analysis or calendar entries.
- Query Planner: The core engine responsible for subgoal grouping and ordering. It is specifically designed to respect "binding patterns," ensuring that reasoners are only queried when their required input parameters are available.
- Reasoning Step Iterators: The system uses an iterator pattern to allow for streaming results and the simultaneous exploration of multiple logical proofs.
Evaluation and Results
The system was evaluated using a suite of 10 complex test queries categorized into Task Fulfillment (TF), Task Setup (TS), and Task Discussion (TD). These tasks represent real-world office scenarios such as equipment procurement, meeting scheduling, and post-meeting analysis.
As shown in Figure 3, the results demonstrate the system's efficiency. Query planning time (QP Time) remained low, ranging from 20ms to approximately 202ms, even for queries involving up to 22 literals and 5 distinct knowledge sources. This suggests that the logical overhead of the Query Manager is negligible compared to the latency of external data retrieval, such as web scraping.
Discussion and Conclusion
The CALO Query Manager represents a shift from passive data retrieval to an active reasoning paradigm. By incorporating binding constraints as first-class citizens and distinguishing between open and closed-world reasoning, the system provides a robust framework for heterogeneous data integration.
Future work focuses on transitioning from static to dynamic query planning, allowing the system to adapt its strategy based on intermediate results. Additionally, the authors identify cost-based optimization and expanded support for complex disjunctive and negative queries as key areas for further development.
Summary updated