For the past months, we have been developing an AI assistant with Bibliothèques Sans Frontières and the Dr. Denis Mukwege Foundation to provide critical legal information to survivor networks of conflict-related sexual violence and the organizations supporting them.

The Challenge

Survivor networks and support organizations face multiple barriers when seeking information:

• Limited access to legal expertise in under-resourced regions

• Complex legal frameworks varying by country and ratification status

• Language barriers across affected regions (English, French, Ukrainian)

• No reliable internet in conflict zones

When one needs information about legal protections or medical care, they need immediate access to verified guidance. The Red Line Initiative’s Guidebook (https://www.endcrsv.org/guidebook/) contains this information, but navigating a comprehensive legal manual is impractical in crisis situations.

Legal Document Intelligence & SLMs

Traditional approaches fail in these conditions. Going through complex international legal documents requires time that is cruelly lacking as well as one has to be sure that no critical element from one of the multiple conventions has been left behind. Paper guides are difficult to navigate in emergencies.

Understanding natural questions – A support worker or survivor can ask "what legal protections exist for survivors in my country" in natural and simplified language, and the system provides relevant legal information from the Redline Guidebook, including ratification dates - without requiring precise legal terminology.

Context engineering for legal accuracy

The knowledge base is limited to verified Red Line Initiative’s guidebook content, ensuring all responses are traceable to authoritative sources. Testing workshops in Ukraine, Niger, and with the University of Cincinnati Law School have revealed implementation challenges specific to humanitarian contexts - particularly maintaining contextualisation for legal accuracy. A comprehensive work has been conducted to enhance and synthetically enrich the initial texts with information “guiding” language models to formulate the fully contextualised, fine-grained response.

Models working online and offline

Two versions: an online assistant for areas with connectivity, and an offline one for conflict zones without reliable internet. The deployment of the offline version has been possible through extensive R&D on pleias’ 350m model.

Supporting multiple languages

Available in English, French, and Ukrainian to serve affected regions including Ukraine, Niger, and other conflict zones.

When information directly impacts survivors' safety and legal rights, the technical requirements become non-negotiable.