A knowledge assistant that never leaves your building.

Everyone has been in the meeting. Someone demos an AI answering questions about a company's own policies, in plain language, with citations, and the room leans in. Then somebody asks the second question: so we upload our contracts where, exactly? And the room leans back out.

Both reactions are correct. The want is real: a company's knowledge lives in contracts, tickets, wikis, case files, and code, and almost none of it is answerable without a human going digging. The stall is real too: that same knowledge is the most sensitive thing the company owns, and "send it to an API" is not a sentence legal will sign. The part almost nobody in the room knows is that the two are no longer in conflict. The entire pipeline that makes documents answerable can run on one machine, in one rack, under one jurisdiction, with nothing leaving it.

What a knowledge assistant actually is

Strip the acronym worship and the mechanism is simple. Your documents get split into passages and encoded into embeddings, numerical fingerprints of meaning. Those fingerprints live in a vector database. When someone asks a question, the question gets encoded the same way, the database finds the passages whose meaning is closest, and the model answers from those passages, citing what it used. The industry calls it retrieval-augmented generation. The important part is what it implies: the model does not memorize your archive. The pipeline consults it, live, on every question.

Which means the pipeline is the sensitive surface. Not just the model: the documents going in, the embeddings made from them, every question anyone asks, every answer that comes back, and the logs of all of it. Decide where the pipeline lives and you have decided where your company's knowledge lives.

The pipeline is the privacy problem

Walk the standard cloud recipe for this feature and count the third parties. Documents are embedded through a hosted embeddings API. The vectors go into a managed vector-database service. Questions route through another API to a hosted model. Logs accumulate at every stop, under every provider's own retention policy. Four external parties, each one an answer you now owe the auditor.

Organizations that would never upload a contract will happily push the same contract through an embeddings API, because "it's just vectors." Embeddings are not anonymous numbers. They are a lossy encoding of the text's meaning, built specifically so that meaning can be searched. A vector database populated from your archive is your archive, transformed. Treat its jurisdiction, its tenancy, and its access controls exactly the way you treat the documents themselves, because functionally it is them.

Once you see the pipeline whole, the single-tenant answer stops being a preference and becomes the design. One machine holds the documents, the embeddings, the model, the questions, the answers, and the logs. The list of third parties who see a token has one entry: nobody. When the auditor asks who else can see this, that is the entire answer, and it is a better afternoon than a data-flow diagram with five logos on it.

What the managed layer means

We size the GPU to the model and the corpus honestly (the pairing logic on the GPU Servers page is the public version of that conversation), deploy and tune the serving stack behind an OpenAI-compatible endpoint, stand up the embeddings model and the vector database beside it, and keep the whole thing running. Your team's integration work is two items long: point the ingestion at your documents, and call the endpoint. The compatibility fights, the driver pinning, the quantization trade-offs happen on our time, which is the entire point of the arrangement. Our managed GPU service sets out the engineering scope.

Who this is actually for

The verticals sort themselves by what they are afraid of. Legal teams, because privilege is existential and a contract cannot touch a third-party API. Healthcare, because a reportable incident is the nightmare and patient context flows through prompts. Fintech and insurance, because the regulator's question has to have an answer with a name on it. Public sector, because sovereignty is increasingly a procurement requirement rather than a preference. And product teams building on open weights, because a knowledge base that trains someone else's model is a moat with a hole in it. Different fears, one architecture.

And the honest boundary, because it is the house habit: if your corpus is small, low-volume, and genuinely not sensitive, a hosted API is often the reasonable answer and we will say so when you ask. This post is for the archive that can't go to a third party. For that archive, the question was never whether the assistant is worth it. It was whether the documents were safe, and the answer is to stop sending them anywhere.

“Embeddings are not anonymous numbers. They are your documents, encoded.”