AI capabilities / Knowledge bases
Knowledge bases over your own private documents
Retrieval augmented generation over your business document library, running entirely inside your federation. Staff query in plain language and get cited answers in seconds, and the documents never leave your walls.
What a private knowledge base does
A private knowledge base lets staff query everything the organization has ever written, in plain English, and get cited answers in seconds. Tax firms query IRS publications. Insurance agencies query policy documents. Clinics query treatment protocols. For any organization with a document library, this is the single highest leverage capability in the federation.
Most organizations have a document library nobody can actually search. PDFs in folders. Procedures in shared documents. Policies in old files. Past memos buried in email. The information exists; finding it is the bottleneck. A private knowledge base indexes all of it into a searchable vector store that lives inside your federation, and the reasoning core writes plain language answers with citations back to the source document.
The interface is a search box that returns answers, not a list of documents. A staff member asks what the procedure is for handling a return after ninety days and gets a paragraph answer citing the actual returns policy memo. Every answer carries a citation, so the person reading it can verify before acting on it.
The architecture, end to end
Source documents run through a chunking pipeline. Each chunk is converted to a high dimensional vector by an embedding model, and the vectors are stored with their source citation attached, inside your own data store. When someone submits a query, the query is embedded and the closest matching chunks are retrieved, typically the top five to twenty. Those chunks, the query, and a system prompt go to the reasoning core, which writes the answer with citations.
The chunking pipeline is where most of the engineering lives. Chunk too large and retrieval becomes coarse. Chunk too small and context is lost. Chunk size and overlap get tuned per document type; a PDF full of tables needs different handling than a memo, and a treatment protocol needs different handling than an IRS publication.
Further reading: the Lewis et al. retrieval augmented generation paper, pgvector, and Qdrant documentation for the self hosted vector store options this capability is built on.
Stack
What the retrieval layer is built from
Document layer
PDF parsing and document structure extraction ahead of the chunking pipeline.
Vector layer
BGE and E5, run locally inside the federation so document content never leaves for an embedding API.
Storage
Self hosted pgvector or Qdrant, holding your citations alongside your vectors.
Generation
The reasoning core already running inside your federation writes the cited answer.
Where this fits
Industries that benefit most
Publication and code retrieval
Retrieval over IRS publications, state codes, and firm policy memos.
InsurancePolicy document retrieval
Cited answers across the full policy library for service teams.
MedicalTreatment protocol retrieval
Clinical staff query indexed guidelines and treatment protocols directly.
EcommerceSemantic site search
Vector retrieval across the product catalog, replacing brittle keyword search.
FAQ
Knowledge base questions
Will the knowledge base ever see PII or protected health information?
That depends on the corpus you index. Because retrieval and the answer model both run inside your federation, sensitive documents are never sent to an outside API to be indexed or read.
How fresh are the answers?
A knowledge base reflects whatever has been indexed. The federation can reindex on a schedule, nightly or weekly, or the moment a document changes.
Can I remove a document and have it disappear from answers?
Yes. Documents are individually removable from the vector store, and the next query will not return content from a removed document.
What about access control across roles?
Documents can be tagged with access groups, and every query is filtered to the requesting user permissions before retrieval runs.
How is this different from a custom GPT pointed at my files?
A custom GPT sends your documents to a third party model, fine for low scale internal use. A federation keeps chunking, embedding, retrieval, and access control inside your own walls, which matters for anything business critical.
Begin
Ready to make your documents searchable?
The readiness assessment covers your document library alongside everything else your federation would need to run.