Private RAG
Chat with thirty years of documents. Inside your walls.
Retrieval-augmented generation is how AI answers from your files instead of hallucinating about them. NetRay builds private RAG systems that run entirely on your infrastructure: permission-aware, citation-first, and grounded in the file shares, wikis, and ERP records you already have.
Why most RAG demos die in production
The demo is easy: embed some PDFs, answer some questions, applause. Production kills it in three ways: chunking that shreds tables and drawings so answers miss the substance; retrieval that ignores permissions so an intern can query the M&A folder; and no citations, so nobody trusts anything and usage quietly dies.
Production-grade private RAG treats all three as first-class engineering: structure-aware parsing for the documents enterprises actually have (scanned PDFs, spreadsheets, engineering drawings, ERP exports), retrieval filtered by your existing ACLs at query time, and a citation for every claim.
Your data sources, connected honestly
The value compounds with coverage. We connect the sources that hold your real institutional knowledge, with the permission model intact for each:
- File shares and SharePoint: decades of specs, procedures, and project folders
- Wikis and ticketing: the tribal knowledge that leaves when people retire
- ERP and PLM extracts: parts, BOMs, routings, quality records (our home turf: SyteLine, LN, M3, NetSuite)
- Email archives and contracts, where policy allows, with tight scoping
The stack, all inside your perimeter
Embedding models, vector store, reranker, and the generation model all run on your hardware. That combination is what makes it private: there is no 'we only send snippets' asterisk, because nothing is sent anywhere. For regulated clients this is the difference between deployable and banned.
This is also exactly what our DataRay product productizes: connect any data source, ask anything, on-prem. Custom builds and DataRay share the same architecture; the engagement tells us which fits.
Measured like a system, not a vibe
We ship every RAG deployment with an evaluation harness: a question set with known-good answers drawn from your corpus, scored for retrieval hit rate, answer faithfulness, and citation accuracy. When you add documents or swap models, you rerun the harness and see the numbers move. That is how RAG stays trustworthy after the launch demo.
Straight answers
How is this different from uploading files to a chatbot?
Three ways: scale (millions of documents, continuously synced, not twenty PDFs per chat), permissions (retrieval respects your ACLs per user), and location (every component runs inside your network, so nothing is uploaded anywhere). It is infrastructure, not a workaround.
What about scanned documents and drawings?
Handled deliberately: OCR with layout awareness for scans, table extraction for spreadsheets and reports, and metadata-first indexing for drawings. Legacy corpora are usually messier than teams expect; our parsing pipeline is built for that mess.
How do you prevent hallucinated answers?
Grounding plus honesty: the model answers only from retrieved passages, cites them inline, and says 'not found in the corpus' when retrieval comes back weak, rather than improvising. The evaluation harness measures faithfulness so regressions are caught, not felt.
How long until our team can query real documents?
A pilot over one corpus typically runs 3 to 5 weeks including the evaluation harness. Full production with multiple sources, SSO, and permission mapping is usually 8 to 12 weeks.
Your documents already know the answer.
Bring a sample of your messiest corpus. We'll show you what grounded, cited answers over it look like, inside your walls.
Scope a private RAG buildFree AI workshops available. A founder reads every message.