← Back to cases · CASE 04 · AI

Enterprise AI
Knowledge Base

Law firm · 5,000+ documents · semantic search + citations · on-premise intranet deployment

5000+Docs indexed
30sSingle search
100%On-premise
Enterprise AI Knowledge Base
01 / Background

5,000 documents in a shared drive. Finding one takes 30 minutes

A 30-person law firm. Years of contract templates, legal opinions and case files sat in shared-drive folders — search by filename, flip by memory; finding one old contract clause took 30 minutes. New hires needed two months just to learn the documents.

The hard constraint was security: client data could not leave the intranet, which ruled out every SaaS knowledge tool. The answer had to be on-premise, with data never leaving the building.

02 / Core Features

Ask a question, get an answer with sources

Auto Document Ingestion

Drag PDF/Word/scans in; OCR handles scanned contracts. No manual filing.

Semantic Search

"What did last year's supplementary agreement say about penalties?" — search by meaning, not filename.

Citation Tracing

Every sentence in an AI answer links to source document and page. One click to the original text.

§

Permission Isolation

Visibility by department and matter; lawyers see only their own case files.

On-premise Deployment

Entire stack on the firm's intranet server — models, vector DB and documents all local.

Audit Trail

Every search and answer logged, meeting internal compliance requirements.

03 / Tech Details

On-premise is the floor; retrieval quality is the brand

Doc parsingPDF/Word parsing + OCR for scans; tables restored structurally; legal clauses chunked with context preserved.
RAG retrievalHybrid vector + keyword ranking, 94% recall on legal documents; answers forced to cite source fragments.
On-premise LLMOpen-source Qwen deployed on the intranet — 16G VRAM server is enough, zero external API dependency.
Vector DBLocal vector store, 500k+ chunks retrieved in milliseconds, incremental updates and versioning.
SecurityEverything on intranet + audit logs + document-level ACL; fully functional without internet.
TuningContinuous tuning with real queries — answer accuracy from 78% to 94% after launch.
04 / Client Success

"One old clause: from half an hour of flipping to one question"

30min→30sSingle search
-50%Onboarding time
5000+Docs indexed
0Data left the intranet
"Finding an old contract clause used to take half an hour. Now one question returns it with the source page. Best of all, everything stays on our own servers — client data has never left the building."
—— Client · Law firm partner

Documents everywhere, search painful?

Knowledge bases, internal Q&A, private AI — all doable. Tell us your document volume and compliance requirements.

Free consultation