Nothing leaves your network
Source files, OCR text, the index, and the entity graph are written only to disk on your own hardware.
Private eDiscovery · For Legal Teams
Search privileged documents without uploading a single page.
Private, on-premise AI document review: hybrid search, an entity knowledge graph, and grounded answers that cite their sources — running on hardware you control. Your privileged and confidential files never leave your network, and no third-party model ever sees them.
Review your case documents with AI that runs on your own computer. Search them by meaning, get answers that quote the exact source, and see a map of everyone involved — while your privileged files never leave your office and no outside service ever sees them.
On-premise
No cloud upload
Privilege-safe by design
Entity graph
Auditable evidence trails
The Privilege Problem
Public AI can waive the very privilege you're protecting.
Courts have started to rule on what happens when privileged material meets consumer AI — and the early answers are not reassuring.
Courts are starting to rule on what happens when you put privileged material into a public AI tool — and the early answers should worry you.
Consumer AI is not privileged.In United States v. Heppner (S.D.N.Y., 2026), a federal court held that exchanges with consumer generative-AI tools like ChatGPT are not protected by attorney-client privilege — and courts have begun finding that routing material through public AI can jeopardize work-product protection. Public platforms may also retain inputs to train their models.
Consumer AI isn't private.In a 2026 federal case (United States v. Heppner), the court said conversations with tools like ChatGPT aren't protected by attorney-client privilege — and using public AI can jeopardize your work-product protection. These tools may also keep what you type to train themselves.
The takeaway is simple: the only way to use AI on privileged material without that exposure is to never send it to a public model in the first place.
The takeaway is simple: the only safe way to use AI on privileged material is to never send it to a public tool in the first place.
This is industry background, not legal advice — consult your own counsel on privilege and your obligations.
The Private Answer
Run the AI where the documents already live.
BDS (Batch Document Search) is on-premise software — an on-premise alternative to the cloud eDiscovery platforms most review runs on. All AI inference and all storage are local; the core pipeline makes no outbound network calls and is closed to the network by default. Nothing about a matter transits a third party. See the deployment & security model →
BDS runs entirely on your own hardware — an on-premise alternative to the cloud eDiscovery tools most review runs on. All the AI and all the storage are local, and the core system makes no internet connections at all — nothing about a matter ever passes through someone else's servers. See how it's deployed and secured →
No third-party model
Text, OCR, embeddings, and entity extraction all run on local models — no external AI service receives your documents or prompts.
Air-gappable
The core ingest → search → graph pipeline can run with no internet access at all.
You control the perimeter
Loopback-closed by default; any access beyond the host is governed by your existing firewall.
Nothing leaves your office
Your files, the text pulled from them, and the map are all saved only on your own computer.
No outside AI
All the AI runs locally — no outside service ever sees your documents or your questions.
Works with no internet
The core system can run on a machine with no internet connection at all.
You hold the keys
It's closed to everything by default; you decide who can reach it, through your own firewall.
More Than a Chatbot Over Your Files
Find what matters — and see how it connects.
"Private" is table stakes. The difference is what you can actually do with a document set once it's in place.
Keeping things private is the easy part. The real difference is what you can do with the documents once they're loaded.
Hybrid search, not keyword-only
Vector (meaning-match) and BM25 (word-match) run together, with a signal on every result showing which methods agreed — so you know how much to trust a row before you open it.
Grounded answers that cite sources
Ask the corpus a question and get an answer built only from retrieved passages — and told plainly when the documents don't support it. Where to look, not what to conclude.
An entity knowledge graph
People, organizations, and other entities pulled from the set, with how they connect and co-occur across documents — the relationships a linear review misses.
Auditable evidence trails
Every result and answer traces back to the source document, so a reviewer can inspect how a lead was produced rather than trust a black box.
Search by meaning, not just words
It searches for what you mean as well as the exact words you type, and tells you which kind of match each result is — so you know how far to trust it.
Answers that quote the source
Ask a question and get an answer built only from what's in the documents, with the passages it used — and a plain "not in here" when the answer isn't.
A map of everyone involved
It pulls out the people and organizations and shows how they connect across the documents — the links a straight read-through misses.
Every lead is traceable
Every result and answer points back to the exact source document, so a reviewer can check how it got there — no black box.
BDS is the private intelligence layer over your documents — it makes a set searchable, mapped, and answerable on your own hardware. It runs alongside your existing review workflow, not as a cloud platform you upload to.
BDS sits over your documents on your own hardware — making them searchable, mapped, and answerable. It works alongside your usual review tools; it isn't a cloud service you upload to.
Proof
Watch it run on a real litigation docket.
We ran BDS on a live sample from the SEC v. Ripple docket — ingestion and OCR, hybrid search with grounded answer synthesis, and the entity graph — all captured locally, at real speed, on real court filings. Not an abstract demo. The full case file — searchable, answered, and mapped — is shown below.
We ran BDS on a live sample from the SEC v. Ripple case — reading it, searching it with cited answers, and mapping who's involved — all recorded live on one computer. Real filings, not a mock-up. The full case file is shown below.
A Living Archive
And it's cumulative. Keep adding documents as the matter develops and they join the existing corpus — so a file that grows over years stays one searchable, permanent archive.
And it keeps growing. Add documents over the years and they join what's already there — so the whole file stays one searchable, permanent archive.
A Real Case, Not a Demo · SEC v. Ripple
We ran BDS on the SEC v. Ripple docket.
The SEC v. Ripple securities-litigation docket is a large public record — real court filings, no synthetic data. Here's what one real run produced.
381
documents
12,397
parties, agencies & assets
121,718
connections found
285
entity profiles
One real matter, shown to illustrate the pipeline — an illustrative case, not a benchmark average.
Real questions, answered from the documents.
Three questions put to the case file and answered by BDS from the source documents — each answer built only from what the record says, with the source shown. Nothing here is hand-written.
What is the SEC's core claim against Ripple?
The SEC alleges XRP is an “investment contract” — a security under the Howey test — so Ripple's sales of it should have been registered. The case turns on whether XRP is a security.
Source: SEC complaint & court filings
Who are the individual defendants?
Ripple CEO Bradley Garlinghouse and co-founder/chairman Chris Larsen, whom the SEC alleged aided Ripple's unregistered securities sales.
Source: SEC complaint
How does the SEC say XRP was sold?
As an unregistered security — offered and marketed to investors to raise capital.
Source: SEC complaint
Grounded answers from hybrid search — where to look in the record, not a verdict on it.
The Map
The people, companies, and connections — mapped.
A readable slice of the entity map from the same run — the Obsidian-style vault BDS builds: each dot is an entity found in the documents, sized by how often it appears; each line joins two that show up together. Tap any dot to drill into it, or use a question above to light up its cast; drag to pan, scroll to zoom.
You’re seeing a readable slice — not the whole graph.The full map from this run holds 12,397 entities and 121,718 connections — far too dense to read on a page. So this shows the strongest 16-entity backbone: the people, organizations and places that come up most, and the ties between them. Drag, zoom, and click to explore — the full graph lives inside the system, on your hardware.
Every dot and line is drawn from the source documents — nothing is invented.
Private AI. On Your Hardware.
Try it on a sample of your own matter.
Send a representative, de-identified sample — one matter, roughly 100–300 documents, scanned PDFs included — and a BDS evaluation returns searchable results, an entity graph, and entity profiles drawn from your own material. Your privileged production stays in your environment; a real deployment runs entirely on-premise.
Send us a sample of your own documents (with sensitive details removed) — one matter, around 100–300 files, scans are fine — and we'll show you the search, the map of who's involved, and short profiles of the key people, built from your own material. Your real production stays with you; a full setup runs entirely on your hardware.
Other practices: Investigations · Due Diligence · Real Estate