BeeHyv built the Sanitation.network website, and the chatbot at its centre runs on Genie, BeeHyv's enterprise RAG (retrieval-augmented generation) platform. The first job was turning a large, varied collection into something the chatbot could search.

Genie processed each PDF page by page, so the text in charts, tables and scanned pages became searchable along with the prose. Videos were transcribed into time-stamped segments, which keeps track of where each passage sits in its recording. A language model then extracted the entities in each passage and the relationships between them, and recorded them in a knowledge graph.

When someone asks a question, Genie combines a search of that knowledge graph with semantic search to find the handful of passages most likely to hold the answer. Those passages go to Claude, using the Opus and Sonnet models, through Anthropic's API.

For Sanitation.network, the answer prompt tells Claude to base its answer on those passages alone. The sources are shown with every answer, so readers can identify and check the original material. When the passages don't cover a question, the prompt also tells it to say that the knowledge base doesn't have the answer yet and to suggest a related question it can answer. For a tool that policymakers may use, a clear "not covered yet" is more useful than a confident guess.

Language support comes from BHASHINI, the Government of India's language platform. When a user types a question in a supported language, BHASHINI translates it into English. Genie finds the relevant passages, and Claude writes the answer. BHASHINI then translates the answer back into the user's language, with the source references still attached. Because every question is translated before retrieval, all of them are searched against the same knowledge base.
Beyond the website, BeeHyv configured Genie, ingested the content, built the knowledge graph, onboarded Dasra's team and supported the public launch. It continues to host, maintain and support the platform.