RAG & EMBEDDINGS

PII-safe RAG, from ingestion to retrieval.

Context-aware ML removes PII across 60+ entity types and 50+ languages before your documents reach the splitter, the embedding model or the index—and filters retrieval output at query time. Runs entirely inside your own environment.

30-minute call with a Limina engineer. No deck, no discovery script.
For AI, ML and platform engineers building retrieval systems on internal data.
IMPACT

70,000

words per second on GPU, batch or async

60+

entity types across 50+ languages

15+

file formats: text, PDF, DOCX, images, audio, DICOM, JSON

0

bytes of your corpus sent to a third party

Why RAG leaks.

Embeddings are one-way

Once text is vectorized the identifier is in the index. Deleting the source row does not remove the vector, and the only real fix is a full re-index.

Retrieval doesn't know who's asking

Permissions gate the document, not the sentence inside it. A correctly authorized query still returns the name, the diagnosis and the account number.

Chunking breaks detection

Split a document into chunks and the pronoun in chunk four has nothing to resolve against. Anything working chunk by chunk cannot link it back to the name.

Clean at ingestion, filtered at retrieval

Blue square with a black plus symbol partially overlaid by a tilted light blue eraser.

Runs before the splitter

De-identification happens at document level, where coreference resolution can still link every mention of one person across the whole file. Then chunk and embed a corpus that never contained the identifier.
Blue square with a black plus symbol partially overlaid by a tilted light blue eraser.

Filters what comes back out

The same engine inspects retrieval output, prompts and agent traffic at query time, at 70,000 words per second. Belt and braces for the content already in your index and anything a tool call returns.
Blue gavel with a warning icon above it symbolizing legal caution or alert.

Reproducible, and inside your VPC

The same input always produces the same output, so an index rebuild is deterministic and diffable. One container in your cloud, on-premises, or fully air-gapped. Nothing is sent to Limina.
CUSTOMER WIN

Providence Health

99.5%+

Accuracy on target PHI entities

3

From evaluation to production

Shipped

An AI-powered physician assistant

The AI was ready. The data wasn't.

Providence's Digital Innovation Group supports 50 hospitals and 800 other care facilities. They set out to build a smart assistant for physicians using EHR data and conversation transcripts. Years of clinical data sat unused because it carried too much PHI to feed into a model, and the project stalled in privacy review.

Limina cleared it.

Providence evaluated a major cloud provider first and walked away when they found the service claimed rights to use the data for its own model improvement. Limina deployed as a container into their production systems on Kubernetes. Setup took three days. PHI never left Providence's environment, and the assistant shipped.

Limina's integration was seamless and exactly what we needed to scrub all the PII out of our datasets.

Wayne Foley
Senior Software Development Manager, Providence

Your corpus never leaves your environment.

Container-based deployment means all processing stays in your VPC, your on-premises infrastructure, or a fully air-gapped environment. No third-party access. No outbound data calls. Ever. All processing runs in memory and nothing is written to disk. ISO/IEC 27001 certified, with an accuracy warranty backed by SwissRe.

CONTACT US

Ready to run it on your pipeline?

A 30-minute call with a Limina engineer or solutions expert. No deck. No discovery script. We'll ask what you're working on, where you're stuck, and whether Limina is the right fit.

What you leave with

  • Where de-identification belongs in your pipeline—ingestion, retrieval, or both
  • What accuracy and throughput to expect on your document types
  • A deployment path for your environment, and a realistic timeline

PII in RAG pipelines, answered

Why redact before embedding rather than filtering at retrieval?

Because embedding is one-way. Once a chunk has been vectorized, the identifier is represented in the index, and deleting the source document does not remove the vector—you re-index. Filtering at retrieval is worth doing as a second layer, and the same engine does it, but it is a control on top of a corpus that still contains the data. Cleaning at ingestion is the only version where the identifier was never there.

Can't we handle this with permissions and metadata filtering?

Those solve a different problem. Entitlements decide who can retrieve which document; they say nothing about what is inside it. A user legitimately entitled to a case file still gets the patient name, the card number and the diagnosis in the retrieved passage. Content-level de-identification is what makes the passage safe to return regardless of who asked.

We already have PII in our vector database. Now what?

There is no way to remove it from existing vectors—that is a property of embeddings, not a limitation of any tool. The path is to de-identify the source corpus and rebuild the index. Limina handles the corpus at 70,000 words per second on GPU, which usually makes the rebuild a shorter job than the discussion about whether to do it. In the meantime, filtering retrieval output limits what surfaces from the current index.

Does it run at query time, or only at ingestion?

Both, from the same container. At ingestion it de-identifies documents before chunking and embedding. At query time it sanitizes prompts, inspects retrieved passages, and covers agent and tool-call traffic—the pattern running in production today at the agent and MCP boundary.