Can we build an internal FAQ bot that stays in the office? Yes. The software is the easy part. Owning the answers is the work.
Short answer. Yes. A local model, an index of your approved documents and a chat front end are enough, and all three can run inside the office network. Open WebUI, for example, describes itself as a self-hosted platform built to run entirely offline, supports Ollama and OpenAI-compatible models, and includes document retrieval with hybrid keyword and vector search. The hard part is not the software. It is deciding who owns each answer, keeping documents current, testing the bot with real questions and trick questions, and sending what it cannot answer to a person. A British Columbia tribunal held Air Canada responsible for what its website chatbot said. An office bot is no different.
01
The parts of an internal FAQ bot
Part
What it does
Open option and note
Chat front end with sign-in
Where staff ask their questions.
Open WebUI is self-hosted, works with Ollama and OpenAI-compatible APIs, and says it can run entirely offline. Its README says (as of October 2026) that the code is under the Open WebUI License, which asks that its branding be preserved, plus earlier licences, so read the terms before a wide roll-out.
Model runner
Serves the language model.
Ollama or vLLM, both of which expose an OpenAI-compatible endpoint.
Document index and retrieval
Finds the passages that answer a question.
Open WebUI lists local retrieval over nine vector databases, with hybrid BM25 and vector search and re-ranking. BGE-M3 is an MIT-licensed multilingual embedding model.
The documents themselves
The facts the bot is allowed to use.
Yours. Each needs an owner, a date and a review cycle.
Feedback and logs
Show where the bot fails.
Thumbs up and down on each answer, plus a record of questions asked.
02
Test it before staff do
Collect 50 questions that staff really asked HR, IT and accounts in the last quarter.
Add 20 questions whose answers are not in any document. The bot must decline all 20 and name who to ask.
Add 10 questions in Hindi and in Hinglish if your staff use them.
Add 5 hostile questions, such as one that asks for another employee’s salary. The bot must refuse.
Mark each answer correct, partly correct or wrong against the source, and agree the pass mark before you start.
Re-run the same set whenever you change the model, the prompt or a large block of documents.
03
What a person must own
One named owner for each document or topic, who approves changes and clears wrong answers.
A review date on every document. A bot quoting a superseded policy confidently is worse than no bot.
A short list of topics the bot must not answer, such as disciplinary matters and individual pay.
A route to a human for everything else it cannot answer.
A plain notice for staff: the bot points to the policy, and the policy document is the authority.
A weekly look at the questions asked and the answers marked wrong.
04
How an office would set this up
Pick one domain, for example leave and expenses, and gather the current documents for it.
Install the front end and model runner on a machine in the office, with sign-in tied to your staff list.
Load the documents with their names and dates, and test retrieval before you test the model.
Instruct the bot to answer only from the passages, quote the clause and say so when the answer is not there.
Pilot with one team for a month, and fix documents rather than prompts wherever you can.
Add the next domain only when the first one meets the pass mark.
05
What ANVE does today
ANVE runs Private GPTs, assistants built on the office’s own material, and its documents function answers questions over the office’s own documents with the source passage shown. A policy helpdesk is one of the agents that can be built with a paid pilot; it is not a shipped product, and the pilot fee is agreed in writing before installation. ANVE serves many named users; roles today are administrator and user, and teams, finer roles and single sign-on are not on ANVE yet. A separate page compares ANVE with Open WebUI.
Private GPTs and the documents function run today in ANVE’s development build. No sealed unit exists yet.
06
Questions people ask
Is Open WebUI free to use?
It is free to download. As of October 2026 its README says the code is under the Open WebUI License, which requires preserving the Open WebUI branding, as well as earlier contributions under their original licences. Read the terms before deploying it widely.
Does it need the internet?
Not once it is installed with its models. Open WebUI’s README describes an offline mode, in which an environment variable stops it trying to download models from the internet.
How many people can use one server?
It depends on the model, the hardware and how people use it. The separate question on sharing one AI server goes into more detail, and a trial with your own staff gives the best answer.
What if the bot gives a wrong answer?
Treat it as a document or retrieval fault first. Find the source passage, fix or retire the document, re-run your test set, and tell the staff who relied on it.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
