Hungarian Public Procurement Authority

Structuring five years of public procurement — and making it searchable in plain language.

A national authority sat on 100,000+ unstructured XML tender documents. We built the AI and data-mining pipeline that structures them, plus a RAG-based legal chatbot that answers case handlers with citations.

Project period

2021 — ongoing (5+ years)

Our role

Design and delivery of the data pipeline, ML+LLM curation workflow and the internal chatbot; retrospective extraction for EU data due diligence.

Context

Public procurement data decides where public money goes. The authority’s archive grew by thousands of XML filings a month, each structured differently, each legally meaningful.

The costly problem

Case handlers spent roughly 35 minutes per research query — opening filings one by one, cross-checking legal references manually, re-keying findings into reports.

Baseline

Measured over 4 weeks with 12 case handlers: 35 min median per query, 11% of answers later corrected after supervisor review, no coverage statistics at all.

Intervention

A hybrid pipeline: deterministic XML mapping where structure exists, ML models where it varies, and an LLM layer for legal-language retrieval. The chatbot answers in Hungarian with citations to the exact filing and paragraph, so every claim is inspectable.

Human responsibilities

  • Case handlers approve every extracted field set before it enters the official register
  • Legal officers review a weekly sample of chatbot answers; failures feed the evaluation set
  • The authority’s IT team owns the deployment — we document and hand over, they operate

Measured outcome

35 → 6 minmedian research time per query, measured over 8 weeks after rollout
100k+documents structured and searchable, full 5-year retrospective coverage
2.4%post-review correction rate, down from 11% at baseline

Limitations

Savings are measured for research and retrieval work, not for legal judgment — that stays with the officers. The EU compliance scope covers the defined data-due-diligence fields, not a blanket certification.

Research queues cleared same-day instead of piling up; case handlers shifted time to substantive legal review. Training took two half-day sessions; adoption tracked weekly and held above 80% after month two.

Next step

Extending the same pipeline to award-decision documents and supplier-history analytics.