Documentation chatbot

Building a documentation chatbot for a WHO environmental surveillance platform taught us more about constraint than capability.

The ES Data Platform (https://es.world) supports polio eradication surveillance across multiple countries. Users work across time zones, languages, and connectivity conditions. Waiting on support isn’t always an option.

So we built a RAG-based assistant on top of the platform’s documentation, using Claude Haiku. A few choices we made early on:

Single-turn only – no conversation history. Each question gets answered on its own. Sounds limiting, but it keeps behavior predictable and costs flat. That matters when you’re serving a global user base that doesn’t have a billing relationship with you. Authentication-gated, tied to existing platform accounts. We know who’s asking, and usage stays within expected bounds.

We log queries – partly for auditability, partly because it’s turned into one of the best signals we have for where the documentation falls short.

Multi-language support is mostly handled by the model itself, which is a big deal when your users span francophone Africa, South Asia, and beyond.

Curious how others are navigating LLM-powered support in constrained environments – spotty connectivity, multilingual users, tight budgets. What trade-offs have you had to make?

#PolioEradication #EnvironmentalSurveillance #RAG #LLM #PublicHealth #GIS #WHO #AI