One AI desk over a ministry's scattered knowledge
Confidential · GovTech
The business said
- “Add staff to compile briefing files”
- “Build another document portal”
- “Index everything behind a keyword search”
- “Keep answering parliamentary questions from memory”
- “One platform that gathers, understands, and answers from every source”
The ministry's answers already existed. They were spread across portals, PDFs, and transcripts that nobody could query in one place.
An AI knowledge platform: automated collection, document intelligence, grounded answers
A federal ministry runs on questions. Parliamentary questions that must be answered precisely, speeches that must cite the right schemes and figures, and briefing notes assembled from documents scattered across departmental portals and archives. The institutional knowledge existed, but it lived in too many places, and every answer meant a manual hunt through them. Keyword search does not help when the question is phrased differently from the document that answers it.
We built a single AI platform over all of it. Automated collectors continuously pull from the relevant public sources: parliamentary question archives, departmental portals, and scheme databases. A document-intelligence layer parses the incoming PDFs and transcripts into structured content, embeddings land in a vector database for semantic retrieval, and a knowledge graph links schemes, places, and programs so related facts stay connected rather than isolated in files.
On top of that foundation sit the tools officials actually touch: smart answers that respond to a question with language-model output grounded in the retrieved sources, and speech management that helps draft and organize public statements from the same evidence base. The stack is React and Redux on the front, Node and Express behind it, PostgreSQL and the vector store for data, open-source language models alongside orchestration frameworks, all containerized and deployed on the cloud.
The shift is from searching to asking. A question that used to mean a day of compiling now returns a sourced draft in moments, with the citations to check. The scattered archive is still scattered on the outside; inside the platform, it behaves like one library that has read itself.
- vector search, retrieval
- parsed to structure, documents
- knowledge graph, connections
- automated scraping, collection
- grounded, cited, answers
