Internal legal RAG search — 71% research time reduction at Dr. Horváth és Társai Ügyvédi Iroda Bt.

The client

Dr. Horváth és Társai Ügyvédi Iroda Bt., a ~60-person Budapest law firm specialising in commercial law and data protection, with 12+ years of Hungarian precedent archive in Word, PDF, and email opinions across internal folders and SharePoint.

The challenge

Finding relevant precedents for new cases took juniors 2-3 hours on average. The firm had addressed similar matters before — they just couldn't find them. 'No time to search, rewrite from scratch' was the unspoken default.

The target

A Hungarian-language searchable RAG system over the full archive that understands legal content, not just keywords. Every answer with citations. Data must stay in the EU. Access via RBAC for active attorneys only.

Why UseAIEasily?

Hungarian-language competence (inflected, abbreviated legal vocabulary), EU-compliant architecture experience (GDPR + attorney-client privilege), and track record of production-ready search with quality metrics — not a generic 'chatbot'.

Architecture

Hybrid search: Word/PDF/HTML/EML ingestion with section-aware chunking, multilingual-e5-large + BGE-m3 combined embeddings for Hungarian accuracy, Qdrant self-hosted in EU region, BM25 + metadata filters (year, practice area, attorney), BGE-reranker, Claude Sonnet 4.6 with citation-bound prompts. RBAC middleware gates every retrieval. Full audit trail exports to external compliance.

Delivery

Discovery (2 weeks), PoC (4 weeks) with 500 docs and 80-question eval suite, Production (6 weeks) with full archive ingest, 3-tier RBAC, desktop + mobile UI for partners, monthly re-indexing, cost caps.

Results (3 months post-launch)

• Average research time: 165 min → 48 min (−71%)
• Precedent retrieval accuracy: 89% top-5 relevance on Hungarian queries
• Junior billable hours: +22% (more capacity per associate)
• Hallucinations: 0 incidents over 3 months (347 queries, all with source citations)
• RBAC violations: 0 (traceable in audit logs)
• Partner satisfaction: 9.2/10 (internal survey)

Lessons

Three pillars: (1) paragraph-level chunking along legal section boundaries (not fixed token), (2) hybrid search (vector + BM25) for precise citation retrieval, (3) re-ranker promoting Hungarian-legal-relevant answers. Claude Sonnet 4.6 handled long Hungarian legal text well — but only after correct chunks were inserted. The 80-question eval suite needed 3 iterations before reaching trustworthy accuracy.

Costs and payback

Build: €92,000 fixed-scope (premium over standard RAG due to compliance audit + 3-tier RBAC). Monthly runtime: ~€900. Payback: 60 active juniors × 4 hours/week saved × €80/hr = ~€19k/mo value. Investment recouped in 5 months.