Human-Centered, AI-Assisted Knowledge Production

Agentic AI for Knowledge Production

This research focuses on designing, developing and validating the agentic AI layer of ReactAI: a modular framework in which an orchestrator coordinates specialised agents to reconstruct operational knowledge from heterogeneous cultural and creative evidence.

The framework combines retrieval-augmented generation (RAG) over knowledge graphs with specialised agents capable of retrieving and interpreting additional evidence. Retrieval can use entity neighbourhoods, controlled vocabularies and previously validated decisions, while web-sourcing agents can retrieve external evidence from repositories, authority files and sector-specific databases.

The architecture is intrinsically multimodal. Agents can operate across textual, visual and audio evidence and combine observations from different modalities when constructing candidate knowledge.

Agents produce candidate claims rather than direct writes to the knowledge graph. Candidate claims are accompanied by supporting evidence, calibrated confidence, machine-readable provenance and human-readable explanations so that their basis can be inspected before they are accepted as operational knowledge.

Claim Taxonomy, Task Formalisation & Dataset Annotation

The research formalises the knowledge-production tasks that specialised agents are expected to perform and develops the corresponding claim taxonomy and annotated datasets.

Initial task families include identity resolution and linking, provenance reconstruction, rights-relevant claim generation, and gap-filling and inconsistency detection.

These tasks provide a common basis for developing, comparing and evaluating specialised agents. Dataset annotation connects source evidence with candidate claims and validated outcomes, enabling systematic experimentation with alternative models, prompts, retrieval strategies and orchestration approaches.

Metadata Agentic Framework

The agentic framework combines orchestration, retrieval and specialised agents within a common knowledge-production environment.

An orchestrator determines which specialised capabilities are required for a particular knowledge-production task and coordinates their execution. Agents can retrieve contextual information from the existing knowledge graph through RAG and, where appropriate, seek additional evidence from external sources.

The framework is designed to support multiple specialised agents rather than assuming that a single model or agent can adequately perform every reconstruction task. Different agents can therefore specialise in particular evidence types, modalities, knowledge domains or claim classes while contributing through a common candidate-claim interface.

The resulting architecture supports experimentation with alternative orchestration strategies, retrieval mechanisms, models and agent configurations while maintaining a consistent separation between computationally generated candidate claims and validated knowledge.

Human-in-the-Loop Validation & Continuous Improvement

Candidate claims are presented to experts through a claim review workspace rather than being written directly into the knowledge graph.

Reviewers can inspect the proposed claim together with its supporting evidence, provenance, confidence and explanation. Their decisions provide the validation required before candidate claims become accepted knowledge.

Human validation also creates structured feedback for improving the agentic system. Expert decisions can be used periodically to improve agents, prompts, orchestration strategies, retrieval policies, exemplars and confidence calibration.

The purpose of this feedback loop is not to remove human validation from consequential knowledge production, but to improve the quality of candidate claims and reduce unnecessary expert effort over time.

Systematic Evaluation, Scalability & Release

The research systematically evaluates the agentic framework and its individual components across knowledge-production tasks, modalities and cultural and creative domains.

Evaluation includes the quality of candidate claims, retrieval and evidence selection, confidence calibration, usefulness of explanations, expert validation effort and the contribution of different agents and orchestration strategies.

Scalability is treated as a research problem rather than only as an engineering requirement. Experiments investigate how the framework performs as the number and diversity of assets, evidence sources, candidate claims and specialised agents increase.

The resulting methods, evaluation results and reusable components provide the basis for releasing and further developing a modular human-centred agentic AI framework for cultural and creative knowledge production.