Representative engagement
A search platform over decades of reports, resolutions and meeting records. It answers questions with citations in the language asked and drafts briefing notes for staff to review.

- Client
- A Geneva-based international organisation
- Sector
- International Organisations & NGOs
- Duration
- 14 weeks, including rollout
- Services
- Data & AI Platforms · Adoption & Enablement
01Challenge
Staff spent a lot of time looking for precedent. What had been agreed, reported or recommended on a topic in previous years. The archive held hundreds of thousands of documents in several official languages, many as scanned PDFs, spread across systems built at different times.
Previous search tools matched keywords in one language and returned long lists. The organisation needed accurate and neutral answers, traceable to the source paragraph. Documents with personal data could not go to external services.
02Approach
We designed the platform with the organisation’s IT and legal teams to fit its data protection framework and privileges. Personal data was classified and filtered at ingestion. The model runs in an environment controlled by the organisation.
Quality was evaluated per language with a test set written by subject-matter staff. It holds 300 questions with known answers and sources. We then trained a first group of 60 staff, built briefing-note templates with them and extended access department by department.
03The system
Documents are parsed, including scans and tables, chunked with their structure preserved and indexed for hybrid search across languages. Answers cite the exact paragraphs used. Briefing notes are drafted from approved templates.
Components
- 01Ingestion pipeline with OCR, layout parsing and personal-data filtering
- 02Cross-lingual hybrid search over the full archive
- 03Answer generation with paragraph-level citations
- 04Briefing note drafting from approved templates
- 05Per-language eval suite written by subject-matter staff
- 06Usage and quality dashboard for the knowledge management team
Stack
- Open-weight and hosted LLMs
- Organisation-controlled cloud
- OCR and layout parsing
- Vector and keyword index
- Eval harness
04Results
A Geneva-based international organisation
- −70%
- Time to locate relevant precedent
- 6
- Languages answered with source citations
- 92%
- Answers rated correct and well-sourced in staff evaluation
- 600+
- Weekly active users after three months
Representative engagement. Client withheld under NDA. Details are anonymised and combined so that no client can be identified; metrics are those measured on the engagement.
Related work
Next step
Tell us which process takes your team the most time. In a thirty-minute call, we can tell you if the sprint is a good start.


