AI Virtual Assistants
A virtual assistant is only as good as the material it reads. Ours are built to answer from your company documents and nothing else: policies, product manuals, price lists, help articles, past answers your staff have approved.
When the documents do not contain the answer, the assistant says so and offers a person. That single behaviour matters more than any choice of model, because a confident wrong answer costs more than no answer.
Typical uses are an internal help desk for staff, a product adviser for customers, and a first line for support that collects the details before an agent takes over.
What the assistant does
Each item below is a separate piece of engineering. We build the ones you need.
Finds the relevant passage
Documents are split into sections and indexed. A question retrieves the few sections most likely to hold the answer, using keyword and meaning-based search together.
Answers with a citation
The reply is written from the retrieved text and links to the page it came from, so the reader can check it.
Refuses when the material is silent
If retrieval returns nothing close enough, the assistant does not guess. It states that it has no information on the point.
Respects permissions
A sales assistant should not read payroll policy. Access rules from your document store are applied at search time, per user.
Passes the case to a person
The visitor can ask for a human at any moment. The transcript travels with the hand-off so nobody has to repeat themselves.
Carries out simple actions
Looking up an order status or booking a call through your API. Anything that changes data or spends money waits for confirmation by a person.
From documents to a working assistant
The slow part is rarely the software. It is deciding which documents are current.
Inventory
We list the sources, mark each as current, outdated or contradictory, and agree who keeps them up to date afterwards.
Test questions
Together we write a set of real questions with the expected answers. This set is the measure for every later change.
Build and tune
Indexing, retrieval, prompt and hand-off rules are adjusted until the test set passes to a level you accept.
Staff trial
Your own people use it first and mark poor answers. Customers see it only after that.
Review
Transcripts are sampled each month. Unanswered questions show which document is missing.
What can go wrong, and the safeguard for each
Language models make mistakes. The design assumes it.
Invented detail
A model may add a plausible fact that is not in the source. Citations, a strict instruction to stay within retrieved text, and sampling of transcripts reduce this. They do not remove it.
Outdated source
The assistant repeats an old price faithfully. Every document carries a review date and an owner.
Manipulative input
A user, or text hidden inside a document, may try to override the instructions. The assistant has no access beyond what that user is allowed, so the damage is bounded.
Personal data in conversation
People type account numbers and health details into chat boxes. Logs are retained for an agreed period and sensitive patterns are masked.
Components we work with
The model is replaceable. The index and the test set are what you keep.
- Models
- Hosted LLM APIsOpen-weight models on your own server
- Search
- PostgreSQL with pgvectorOpenSearchHybrid keyword and vector ranking
- Document sources
- PDF and Word filesWeb pagesSharePointGoogle DriveHelp desk articles
- Channels
- Website widgetMicrosoft TeamsSlackWhatsApp
- Application
- PythonNode.jsPHP
Is an assistant the right tool?
It tends to work when
- The same questions arrive again and again, and the answers are written down somewhere.
- Someone is responsible for keeping the source documents correct.
- Staff are available to take over a conversation during business hours.
- You are willing to read transcripts and improve it after launch.
It tends to disappoint when
- The knowledge lives in the heads of two senior people and nowhere else.
- Answers depend on legal, medical or financial judgement about an individual case.
- The aim is to stop customers reaching a human.
- Documents contradict each other and nobody can say which is right.
Practical points
Which model do you use?
The choice follows your constraints: where data may be processed, cost per conversation, and language coverage. We compare two or three on your test questions.
Can it run without sending data outside our network?
Yes, with an open-weight model on hardware you control. Answer quality and hardware cost must be weighed against a hosted service.
How is it kept up to date?
Sources are re-indexed on a schedule or when a file changes. A page removed from the source disappears from the answers.
What does it cost to run?
Hosted models charge by the amount of text processed. We estimate monthly usage from your expected volume and set a spending limit on the account.
Does it speak languages other than English?
Current models handle many languages, with uneven quality. Hindi and Gujarati can be tested by our own team. Other languages need a reviewer on your side.
Send us ten questions people keep asking
Tell us where the answers are kept today and who asks. We will say whether an assistant is sensible for that material. Expect to hear back within one business day.