An index of your own documents, a model endpoint in a location you choose, and answers that cite the page they came from.
Anyone can rent a model by the token. What you cannot rent is a decision about where your contracts sit while a model reads them. There are two versions of that decision, they cost different amounts, and they are not the same product. On your own hardware, your documents never reach a model vendor, because there is nobody to send them to. In a cloud account, the endpoint runs under your name and your vendor's no-training terms govern it. We will show you which vendor, which terms, and which version you are on before anything is installed.
That is the question owners are actually asking. The plain answer is that you would be handing your paperwork to somebody else's servers and taking their word for the rest. If that does not sit right, the alternative is keeping the index and the endpoint on equipment you control.
Documents in, answers out, and a perimeter around everything in between. Where the perimeter falls is the decision you are actually making, and it is the difference between the two prices below.
The index lives where you decide. A public chatbot answers from what it absorbed in training. This answers from what you handed it. Your documents are broken into passages and indexed. When somebody asks a question, the system finds the passages that match, and the model writes from those, with the file and the page attached.
Every deployment is scoped to include document ingest, the search index, a chat endpoint, a citation on every answer, access logging, refusal rules for questions outside the indexed set, and backups of the index. That is the standard we build to and it goes in your written scope. Nobody is running it in production yet, so it is a scope rather than a track record.
These are the three we would build a first deployment against, and the three we would test it on before you rely on it.
"Which of our subcontractor agreements carry a liquidated damages clause."
It has to find every agreement rather than the first three, and show you the clause it is reading.
"What did we quote this customer in 2023, and what did the job actually cost."
Two documents from two systems, joined on a company name spelled two ways.
"Draft the submittal for section 09 21 16 using our standard language."
Retrieve the section, retrieve the standard language, and refuse when the corpus does not contain them.
None of this has been delivered to a client. If you are in construction, treat the pilot as the proof and this page as a description of the approach.
It answers from the passages it retrieved, and every answer shows you which ones. That is not the same as never being wrong. Retrieval narrows what it can draw on. It does not make invention impossible, and a model can still write a sentence the passages do not support. The citation is what lets you check in five seconds. Read it before you act on the answer.
A bad scan produces a bad answer, and OCR quality sets the ceiling on everything above it. Arithmetic across scanned tables needs a checking step. It does not replace a lawyer or an estimator, and it knows nothing outside the corpus you gave it.
Where the files sit, who can reach them, what is logged, and how long backups and logs are kept.
We hold no HIPAA, SOC 2, PCI or ISO certification and will not imply otherwise. If your industry requires certified handling, say so on the first call and we will tell you what that would take and whether we are the right firm to do it.
Access is scoped per person or per group, and which documents each person can retrieve is agreed and tested with you before anyone else gets a login. Pick the test cases yourself. We would rather you try to break it than take our word for it.
Removing a document removes its passages from the live index, and answers stop citing it. Index backups and access logs are a separate question, and we agree the retention window for both in writing before ingest. "Delete" means nothing until somebody has said how far back it reaches.
You hold title to any hardware from the day it is invoiced to you. The index is yours, the ingest scripts are yours, and the documents were never anywhere else. On the day you cancel you receive the runbook, the credentials, and the index in a portable format. On your own hardware it keeps answering after we are gone, because the components are standard and none of them are ours. In a cloud account it keeps answering as long as you keep paying the vendor.
Hub City Web is a trade name of NapCarr LLC, in Lubbock, Texas. An NDA is available before the first document is shared, and we will sign yours.
Two individuals can technically read your data: Oscar, and the contracted engineer named in your scope. On your own hardware, nobody else. In a cloud account, the vendor you selected, under your account. If we find an incident, you hear from us as soon as we know, with what we know then, and again when we know more.
Two steps, both priced.
Document assessment, $400. We take a sample of up to two hundred of your actual files, run the ingest, and show you what came out legible and what did not. It comes off the pilot price if you go on.
Pilot, $2,500. One corpus, one endpoint, and a written list of twenty questions it has to answer correctly, agreed before we start. The schedule is dated in the scope after the assessment. If it fails the list, nothing further is owed and you keep the index and the ingest scripts.
After the pilot it is month to month at the rung that fits, thirty days' notice either way. Nobody is running this in production yet, and the first three engagements carry that structure for that reason.
One document set, one endpoint, up to three people asking. Runs in a cloud account you own, and the compute bills to you.
Up to twenty-five people. Documents ingested as they land in a folder or an inbox, without anyone uploading them. Access rules per person, so the pricing sheets answer for the two people who set them and nobody else.
Your own hardware, in your building, or an isolated cluster only you can reach. No seat limit. Search spans departments. Replication and recovery for the index are scoped with you: say what you need back and how fast, and we price it.
Below about five thousand documents, with two or three people asking, a folder structure and a naming convention beat this and cost nothing. It starts to pay when the documents run into the tens of thousands, when more than about five people need answers out of them, and when the person who knows where everything is has become the bottleneck. If that is not you yet, we will tell you to wait and spend the money on the listing.
Hub City Web is Oscar Napoles, a solution architect. An independent engineer is contracted per project for the infrastructure work: private and on-premise deployments, local model endpoints, ingest pipelines, vector search. He is a contractor rather than staff, and his availability is confirmed in writing before we put a date on anything.
The stack is standard open components. There is no Hub City Web software in the middle of it, and the runbook and credentials are in your possession from the first month. Another competent engineer picks it up from that document.
$500 to $2,000 a month covers the build, the setup and keeping it running. Compute and storage bill to your own account or run on your own machine. We size that number with you and show it to you before you commit. If you are comparing it to something, compare it to the hour a week somebody spends looking for the version of the document that matters.
The assessment is $400, the pilot is $2,500 against twenty questions you write, and nothing further is owed if it fails the list.