Multimodal querying system for UN-mission analysts
RAG-based search over text and audio knowledge bases, so analysts in conflict regions can surface narratives from hours of radio data in minutes.
The repository is private to protect partner data. Drop us a note describing your use case and we'll get back to you.
Walkthrough.
What we built and how.
“Accelerating Narrative Identification in Conflict Regions with Artificial Intelligence” addresses the complex, multi-stage process of identifying misinformation and misleading narratives from radio data for United Nations missions in conflict regions. Human analysts typically spend hours manually sifting through transcripts, summarising findings, and validating sources. To accelerate this, the team developed a retrieval-augmented generation (RAG) based search system that retrieves the most relevant transcripts to an analyst's query.
The system not only provides answers but also presents context and links to relevant source files, aiming to reduce analysis time from hours to minutes without compromising output quality. The design emphasises a human-centred AI approach, ensuring analysts can easily get answers with pre- and post-context, and verify information for their reports.
The project explored various retrieval modules, determining that a GraphRAG-based system with a multi-query retriever significantly improves upon vanilla RAG methods, especially for complex or vague queries. This approach aims to optimise and speed up information retrieval crucial for UN missions in fast-paced information environments.
