Layer 01
Data
Reading, OCR, layout, classification, splitting and extraction, returning typed records with confidence per field and nulls where the document is silent.
Built with structured extraction and local OCR.
Private Document Intelligence
Document AI that never leaves your walls.
AI that understands, extracts, searches, and acts on your documents, entirely inside infrastructure you control: the category, defined.
Stays localdocuments never leave
Citedanswers carry their page
Reviewedhuman checkpoints on changes
Air-gappedruns fully disconnected
Private Document Intelligence is the use of AI inside infrastructure the organization controls to understand, extract, search, transform and act on the information held in its documents, without sending source material to an external service.
Not a product claim: a category definition. The rest of this page states what belongs to the category, and what any system must prove to claim it.
From documents to data, knowledge and action.
Most document projects stop at extraction. Durable value needs all three layers: a field is only useful if something can act on it.
Layer 02
Knowledge
Indexing, lexical and semantic retrieval, and grounded answers that carry the document, the page and the passage, refusing when the corpus is silent.
Built with RAG chat and document search.
Layer 03
Action
Conversion, redaction, archival formats and tools an agent can call, with human review in front of irreversible changes. The output is an artifact you can open and verify.
Built with Smart Redaction and the PDF toolkit.
The problems it exists to solve.
Each one is a scenario page: the pain, the pipeline that removes it, and the capabilities every step uses.
Invoices arrive as scans and someone retypes them
Invoice & document extraction
Policies hold answers nobody can find
Private company knowledge
Sensitive data hides in outbound documents
PII review & redaction
Review queues grow faster than people clear them
Multi-agent document review
Regulated content must be watched continuously
Compliance monitoring
Assistants forget everything between sessions
Assistants with memory
Document workflows stall at every manual step
Agentic document workflows
Staff spend their day asking documents questions
AI assistants on documents
The documents worth automating cannot leave.
Contracts, claims, records, personnel files, scanned archives: the highest-volume document work sits exactly where an external service is blocked.
Control
Residency you can point at
Files, indexes, embeddings and inference sit on machines you name: a claim your auditors can verify, not a compliance promise.
Trust CenterCost
Planned, not metered
Per-token pricing scales with the thing you automated. Owned compute replaces the meter with capacity planning you control.
Cost & PerformanceContinuity
Works when the link does not
Air-gapped sites, field deployments, regulated networks, outages. A process that stops with an external API was never automated.
Edge & OfflineThe bar a system must meet to claim the category.
Five requirements separate Private Document Intelligence from a model behind a proxy. LM-Kit is built against all five.
Local by architecture
Processing, indexes and models run where the documents live
Verifiable by output
Per-field confidence on data, page citations on answers
Governed disclosure
Fully local, or an external assistant sees only the results you allow
Human checkpoints
Review in front of uncertain, sensitive or irreversible changes
Complete to the outcome
From OCR to acted-on result, not a parts kit
What the neighbouring categories optimize for.
All reasonable tools. The useful question is which constraint you are actually up against.
Compared to
Intelligent document processing
Optimizes extraction on known document types, as a managed cloud service. This category delivers the same structured output where documents cannot leave, on schemas you define.
See it on invoicesCompared to
Cloud document AI
Immediate capability and continuous updates, with the file opened elsewhere and a meter on volume. Here the file never leaves and the cost curve is yours.
Local vs CloudCompared to
RAG frameworks
Primitives for engineers assembling a bespoke pipeline. This category arrives working: parsing, retrieval and citations already joined, and maintained as one engine.
LM-Kit.NET vs LangChainCompared to
Private AI infrastructure
Runs models on your hardware: serving, scheduling, acceleration. That is the prerequisite, not the outcome. The work is what happens to the document once a model is available.
Local InferenceA layer beside your systems of record, not a replacement for them. Content management keeps ownership, retention and access; this layer understands, structures, retrieves and acts, then hands the result back.
It meets your stack where it is: in-process .NET, REST, or governed tools for external assistants. See the integrations.
You probably need this when several are true.
One of these alone is usually solvable another way. Three or four together is what the category exists for.
Source documents cannot be sent to a hosted AI service
Document volume makes per-page or per-token billing material
Output has to be structured for another system, not read by a person
Answers need to cite the document and the page they came from
A person must review anything uncertain, sensitive or irreversible
AI assistants need controlled access, without handing them the filesystem
The same capability has to run as a shared service and inside a shipped product
See it working