PDF Diff
PDF Diff embeds every page of two PDFs with an open-weight model running on the user’s Mac or PC, then matches pages by meaning when page numbers shift. Documents never leave the computer.
Answers from your own documents, on infrastructure you control.
Engagement-led · We start with a discovery call
Retrieval-augmented generation, or RAG, finds the passages in your documents that answer a question, then has a model answer from them and show its sources. We build it on your servers or private cloud, so the documents and the index stay with you.
Find the passage, then answer from it.
On its own, a language model answers from what it learned in training. RAG hands it the relevant pages from your documents first, so the answer can point back to them.
-
Index your documents
Files are split into passages and turned into embeddings by a model running on your hardware. The index lives next to the documents.
-
Retrieve
A question is embedded the same way and matched against the index, often alongside keyword search, to find the passages most likely to answer it.
-
Answer with sources
A language model answers from those passages and cites them, so a reader can open the source and check it.
-
Measure
We test retrieval and answers on questions from your own staff before launch, and rerun them before every change.
Where each piece runs.
Private RAG is a set of choices. This is the default we design for, and the one choice that sends text outside.
| Piece | Runs on | What leaves your environment |
|---|---|---|
| Documents and index | Your servers or private cloud | Nothing. |
| Embedding model | Your hardware, as in PDF Diff and Audio Search | Nothing. The model is downloaded once. |
| Answering model | An open-weight model on your GPUs, or a commercial model you choose | Nothing with a local model. With a commercial model, the question and the retrieved passages go to that provider. |
| Logs and test results | Your infrastructure | Nothing. |
Who can retrieve what is a design decision we settle with you in discovery, usually by following the permissions your document system already has. If you need the GPUs, we build those too.
Built on retrieval we already ship.
The retrieval half of RAG is already running in our products and in systems Edgett has built.
Audio Search
Audio Search splits long recordings into short passages, stores their embeddings in a local database, and ranks matches with exact similarity search. Text and audio queries run on-device.
Blazor Chat
Blazor Chat (opens in a new tab), the open source app Edgett created in 2023, chats with local language models on .NET 8 and answers with document retrieval.
SEC filings
For our customer Ebullience, Palm Hill built a platform that extracts executive compensation from SEC filings using retrieval and multi-step extraction. Customer work: Ebullience
The embedding models behind the first two are listed under open weights we already ship.
The documents people search every day.
Legal
Matter files, precedent, and firm policies.
Ask a question across a matter’s documents and get the passages that answer it, with the files, the index, and the model on firm infrastructure.
Healthcare
Policies, procedures, and operations documents.
Staff find the current policy or procedure and see where the answer came from. Deployments that touch clinical documents are designed with HIPAA requirements in mind and reviewed with your compliance team first.
Finance
Filings, credit policies, and procedure manuals.
Answers from the documents your teams already rely on, with a record of who asked what and which passages the answer used.
Operations
Contracts, manuals, claims, and archives.
Older archives often need OCR and page classification before they can be searched. We have built both into document review since 2012.
Where NoPilot and WolfPackAI fit.
Private RAG is something we build with you. Two of our products can be part of it.
- NoPilot is a Copilot-style assistant you host. Its documents are stored and indexed on your infrastructure. Early access.
- WolfPackAI routes the answering model through one gateway, local or commercial, with OpenTelemetry tracing. Alpha, with pilots available.
Connectors and file types
Which document systems and file types a deployment connects to is settled in discovery, based on where your documents live today.
Agents that use it
A retrieval index is also a tool an AI agent can call, for example to find the policy that applies before it drafts a reply.
Questions about private RAG
What is private RAG?
Retrieval-augmented generation where the documents, the search index, and ideally the models run on infrastructure you control. The model answers from passages found in your documents and shows where they came from.
Does any of our data leave our network?
Not with local models. The documents, index, embedding model, and logs stay on your servers or private cloud. If you choose a commercial model for answers, the question and the retrieved passages go to that provider, and we show you exactly where that happens.
How is this different from uploading files to a chatbot?
The files never go to a vendor’s cloud, the index is yours, and you choose the model. You can also measure answer quality on your own questions and rerun those tests before every change.
How do you know the answers are right?
We build a test set from questions your staff actually ask, check that the right passages come back and that answers stick to them, and rerun it before every model or prompt change. How we test.
Can it work with scanned documents?
Yes, after OCR. Image-only scans need a text layer before they can be indexed, and we add that step when an archive needs it.
Do NoPilot or WolfPackAI include RAG?
NoPilot stores and indexes documents on your infrastructure, and WolfPackAI routes and traces the model calls. The retrieval pipeline itself is scoped and built with you, so it fits where your documents already live.
Bring a folder of documents and ten real questions.
Redacted or sample documents are fine. We will show you what retrieval finds, where the answers come from, and what it would take to run on your infrastructure.