From Prompts to Context: An Ontology-Driven Framework for Human-Generative AI Collaboration

Technical Summary: From Prompts to Context: An Ontology-Driven Framework for Human-Generative AI Collaboration

Problem Statement
Current collaborations between humans and Generative AI (GenAI) often suffer from a lack of contextual explicitness. Interactions typically begin with a short, ad hoc prompt and conclude with an opaque output, leaving implicit critical information regarding the agents involved, the specific tasks pursued, the resources utilized, and the constraints governing the process. This "context drift" hinders trust, traceability, and accountability, particularly in information-intensive workflows such as software development, search, and profile management. The opacity arises not only from the underlying neural models but from the broader system-level interactions where prompts, retrieved context, intermediate steps, and human interventions are not formally linked. Existing collaboration ontologies largely focus on human-human interactions or lack the specific granularity required to model dynamic GenAI roles, leading to a gap in standardized methods for capturing the "who, what, how, and why" of Human-GenAI collaboration.

Methodology
The paper proposes "From Prompts to Context," a framework centered on the Contextual Collaboration AI Ontology (CCAI). The methodology involves three primary components:

1. The CCAI Ontology: Developed using the SAMOD methodology and modular "modelet" design, CCAI provides a shared, machine-interpretable vocabulary. It is built upon foundational ontologies like PROV-O and FOAF, extended with domain-specific classes. Key entities include:

   • Agents: Differentiating between ccai:HumanCollaborator, ccai:GenerativeAIAgent, and ccai:AgentGroup.
   • Roles & Competences: Defining specific responsibilities (e.g., ccai:ToolRole, ccai:AIPartnerRole) and capabilities.
   • Tasks & Processes: Modeling work units (ccai:Task) and workflows (ccai:CollaborationProcess) with dependencies.
   • Context & Resources: Capturing temporal, spatial, and domain settings (ccai:CollaborationContext) and tools/datasets (ccai:CollaborationResource).
   • Constraints: Explicitly modeling ethical guidelines (ccai:EthicalConstraint).

2. SPARQL-Driven Prompt Construction: The framework operationalizes the ontology by using SPARQL queries to retrieve task-relevant semantic information (context, resources, roles, constraints) from the knowledge base. This retrieved data is injected into a structured prompt template, merging human intent with explicit ontological bindings. This shifts prompt engineering from free-form elaboration to a structured, context-complete request.

3. Provenance-Linked Artifact Generation: Outputs generated by the AI are linked back to the originating task, agents, and context using provenance relations (e.g., prov:wasGeneratedBy, prov:wasAttributedTo). This creates a queryable collaboration trace, transforming ephemeral interactions into auditable records.

Key Contributions
The paper makes three specific contributions:

1. A Contextual Ontology: The CCAI ontology makes collaboration semantics and context "first-class" citizens in Human-GenAI work, formally defining roles, tasks, resources, and constraints.
2. An Instrumentation Approach: A workflow that integrates SPARQL-derived context into prompting and review processes, producing structured, auditable collaboration traces rather than relying on implicit memory or unstructured logs.
3. A Representative Case Study: An in-depth application of the framework within a software development team building a "View & Update Competency Profiles" feature for a competency-based education system. This study demonstrates the framework's utility across requirements analysis, design, implementation, and testing phases.

Results
The case study provides bounded empirical evidence comparing a "prompt-only" (PO) condition against a "CCAI-backed" (CB) condition:

• Contextual Explicitness: In the CB condition, all contextual categories (context, resources, responsibilities, constraints) were explicitly represented in the prompt (4/4 categories), whereas the PO condition left them implicit (0/4).
• Traceability: The CB condition preserved structured provenance paths for all 12 instantiated tasks, linking artifacts directly to tasks and agents. The PO condition frequently omitted task-linked contextual items (86/95 items omitted).
• Operational Impact: The framework successfully reduced contextual ambiguity and enabled the reconstruction of decision trails. For example, SPARQL queries could instantly retrieve the specific resources (e.g., ccai:CompetencyDB), roles, and agents associated with a task, which were then used to assemble a precise prompt.
• Limitations Observed: The study noted that the effectiveness of the framework depends on the timeliness of the ontology's knowledge base. If resource descriptions in the ontology become stale (e.g., API changes not immediately reflected), the generated prompt context becomes partially incorrect, requiring human intervention. Thus, ontology curation becomes an explicit operational task.

Significance and Claims
The authors position this work as a step toward more transparent and accountable Human-GenAI practices. They explicitly state that they do not claim efficiency or productivity gains; rather, the significance lies in:

• Enhanced Transparency: Making the collaborative context explicit allows stakeholders to understand the rationale behind AI outputs.
• Improved Traceability: Providing auditable chains from user stories to generated code and commits, facilitating better oversight.
• Shared Understanding: Serving as a "boundary object" that aligns human and AI agents on roles, constraints, and goals, reducing the risk of context drift.

The paper concludes by outlining design principles for future systems: context must be explicit and inspectable; provenance should be integrated by default; and the ontology must serve as a shared communication interface, not just a backend schema. The authors acknowledge that while the framework improves traceability, it does not eliminate the probabilistic nature of GenAI or the risk of domain-inadequate suggestions, necessitating continued human review and validation.