Late last year, the compliance team at a regional bank ran an internal tally. Across shared drives, email attachments, and outsourced systems, their policies, regulatory bulletins, and due-diligence reports added up to more than two million documents. They wanted something that sounded ordinary: let staff ask questions in plain language and have the system return answers with citations. Three things stopped them. Data could not leave the intranet. Document formats were a mess, with scans and PDFs making up the bulk. And the real fear — a model that sounds confident while being wrong, where every sentence has to be traceable back to a source.
This is exactly the problem UglyPear Data has been chipping away at for years. What we build is a high-performance document compression and retrieval-augmented generation (RAG) data-engineering foundation, anchored to one principle: fully private deployment, data never leaves the internal network. Let me break down what it actually takes for a financial institution to stand up a private knowledge base that people trust.
Why finance both wants and fears a knowledge base
A large share of knowledge workers' time disappears into looking for things. An IDC survey of 1,200 information workers found employees spend an average of 4.5 hours a week just hunting for documents. In a financial setting that number tends to climb. There are more policies, more versions, more cross-department searches, and half a day can vanish simply confirming which version is the current, valid one.
The bigger problem is how sensitive the data is. IBM's 2025 Cost of a Data Breach report notes that among stolen data, customer personally identifiable information (PII) made up 53 percent, employee PII 37 percent, and intellectual property 33 percent. And for intellectual property, the cost per lost record can run as high as 178 dollars. For a bank, brokerage, or insurer, handing batches of contracts, client files, and research reports to a public-cloud model is close to leaving the vault door open.
The data keeps growing. Komprise's 2026 report states that 74 percent of large enterprises store more than 5 PB of unstructured data, 40 percent exceed 10 PB, and 85 percent of IT leaders expect storage spending to keep rising through 2026. An institution's archives, imaging, and logs are exactly the kind of data that takes up space without being touched often, which quietly inflates the bill.
Why compression is a prerequisite, not an afterthought
A lot of people assume RAG just means dumping documents into a vector store and asking questions. The reality is less tidy. Raw documents fed to a model — large, noisy, full of boilerplate — blow up both token cost and retrieval noise. The answers get diluted, the bill climbs, and the retrieval layer starts returning the wrong chunks. Our approach runs a preprocessing pass before anything enters the store: PDFs, images, video, audio, Office files, OFD, scans, and other common formats get handled uniformly, compressed where it helps, desensitized where needed, and structured where it matters. In our public benchmarks, a 443-page OFD document finishes end to end in about 124 seconds, and image assets compressed before being sent to GPT-4o cut token cost by roughly 85.9 percent. Volume typically drops 60 to 80 percent, which means the same retrieval budget covers more material and the same answer draws on cleaner context.
Recently we pushed the private-knowledge-base line a step further toward something deliverable and verifiable. A minimum viable acceptance system has passed end-to-end testing, covering the full chain from upload and incremental ingestion to source-traced multi-path retrieval, citation-constrained question answering, and usage metering. It can be deployed with a single command as a container inside the customer's intranet, running fully offline. For a financial institution that means standing up an in-house knowledge base where no data is sent to any external service, and every answer carries its source. After optimization, retrieval and answering that once took over a minute now respond in under a second — fast enough that people actually use it instead of falling back to emailing a colleague.
Three routes, and how they really differ
Financial institutions usually face three routes to a knowledge base, and the real differences sit in data sovereignty and document preprocessing.
| Dimension | Build from scratch | Public-cloud managed | UglyPear private RAG foundation |
|---|---|---|---|
| Data leaves the intranet? | Controllable, but you build the compute | Usually uploaded to the cloud | Never leaves the intranet |
| Compliance and audit | Your responsibility | Depends on vendor certifications | Private deployment; answers traceable and auditable |
| Document preprocessing (parse, desensitize, compress) | Self-built, heavy lift | Partial support | Built in, covers 8 common formats |
| Time to launch | Long | Short | Moderate, containerized delivery |
| Best fit | Depends on team | Elastic scaling | Mid-to-large institutions, intranet scenarios |
For institutions with high data-sovereignty requirements, messy document types, and a wish to see results soon, a private RAG foundation is usually the steadier choice. It solves the hardest step — getting documents cleanly into the model — first, instead of treating it as someone else's problem.
A few practical reminders
Handle sensitive information up front. Our foundation ships 13 categories of AI-driven sensitive-information detection, desensitizing regions such as faces, IDs, and stamps so fields that should never reach a model stay out of the context window. Don't expect to finish in one move. Start with one high-value document class, like regulatory bulletins or product manuals, prove it works, then expand. And write "every answer must cite its source" into the acceptance criteria; an answer without traceability is simply not trustworthy in finance, where a wrong citation can mean a wrong decision.
Think about the retrieval budget, not just the model. Compressing before ingestion is what makes a large, heterogeneous document set affordable to query repeatedly. The token savings compound every time someone asks, which is why a one-time preprocessing investment keeps paying off.
FAQ
Q: With a RAG knowledge base, is it true the data never leaves the intranet?
A: Under a private deployment, everything from document parsing, compression, and vectorization to answer generation runs on the customer's own servers or internal network. No external cloud service is involved, so data sovereignty stays with the institution.
Q: Can difficult formats like scans and OFD be used directly?
A: Yes. The foundation covers eight common formats — PDF, images, video, audio, Office, OFD, scans, and more — and can parse and compress scanned and OFD layout files before ingestion.
Q: Does compression destroy key content?
A: Compression follows a usability-first principle. In public benchmarks a 443-page OFD finishes in about 124 seconds end to end, with volume down 60 to 80 percent while readability is preserved, and stamps and layout get dedicated protection.
Q: Is a private deployment more expensive than calling a public-cloud model?
A: In the short term a private deployment needs your own compute, but what it saves long term is token cost and compliance risk. Compressing documents before ingestion sharply lowers token cost (image assets roughly 85.9 percent in our public benchmarks), and it removes the potential loss from data leakage.
Q: Can smaller teams use it?
A: The private foundation fits mid-to-large institutions' intranet scenarios best. Individuals or small teams can start with UglyPear Data's desktop product, SmartSlim, to handle local document compression and organization.
Try it now
If your institution is evaluating a private knowledge base, these are good places to start:
- Product overview: SmartSlim and MagiFiler
- RAG data-engineering foundation: UglyPear Data
- More engineering practice and benchmarks: UglyPear Data blog