FAQ

The questions your compliance team will ask first

Answered plainly, because the answers are the product. If something here isn't covered, bring it to the demo. Architecture questions are the ones we most want in the room.

Book a demo Security overview

What it can do

What file types can it read?: PDFs, Word documents and spreadsheets, plus scans and photographed pages that were never searchable text. Spreadsheets keep their rows and columns, so a figure stays attached to the column it belonged to.
Can it do calculations?: It reads spreadsheets and can compare, aggregate and summarise what is in them, citing the rows it used so the arithmetic is checkable. It is not a replacement for your finance system: where a number has to be authoritative, it shows you the source so a person confirms it.
Does it have internet access?: Only when you switch it on. There is an Internet toggle on every question, off by default. Your documents never travel out through it. It only allows an answer to reach outside when you decide a particular question warrants it.
Can it make charts?: Yes. Ask for a visual and it draws one from the figures it has just cited, so the bars are your own data. The chart carries its own source marker and exports with the document.
Can it generate documents?: Yes, and it exports them. A comparison, a roll-up of findings or an onboarding brief comes out as PDF, Word or Excel, assembled from your own material rather than written from a blank page.
What language does it answer in?: The language you asked in. Ask in Romanian and the answer comes back in Romanian, cited to documents written in whatever language they happen to be written in.
What happens when it doesn't know?: It says so, and says why. In a real answer from a working install, it declined to compare two months because those files held attendance codes and nothing about workload. It then listed exactly what data would be needed. A tool that answers everything is one you have to double-check every time.
Is there a dashboard? How do I administer it?: Four admin areas inside the same app: Users (accounts and roles), Access groups (which groups reach which files), Configurations (model and behaviour) and Usage, which shows consumption per person and per team. Every answer records its response time and token count, so the bill is something you look up rather than something that arrives.
Can I scope a question to part of the library?: Yes. Every question carries a group filter, so you can ask across all groups or narrow it to one first. There is also a Deep reasoning toggle for questions that need more work.

Data & residency

Does any of our data leave our environment? What exactly reaches the model?: Nothing leaves your own infrastructure, and no third-party AI vendor ever sees your documents. The vector store and all document text, OCR, parsing, chunking, lexical search, spreadsheet recalculation and web page reading run as containers inside your deployment, with no exceptions. The model runs in your infrastructure too, wherever you choose: AWS Bedrock in your own account, Ollama on your own servers, or any other infrastructure that can serve it. It receives exactly two things, and nothing is sent anywhere else: passage text at indexing time, to build the search index, and the relevant passages the person may read plus the question at answer time. Your content is never used for training. Exactly what reaches the model.
Where is the data stored?: In your own infrastructure, in the region you deploy into. For EU data residency that is an EU region, and the model runs there too, on AWS Bedrock in your own account, on Ollama on your own servers, or on any other infrastructure you choose, so documents, processing and inference all stay where you put them. Residency follows from where the software was installed, so it is not a setting anyone has to get right.
Is our content used to train models?: Never. Your content is not used to train our models or a model provider's. If the model runs on AWS Bedrock in your own account, Bedrock does not retain your data or train on it; with Ollama or other self-hosted models on your servers, nothing leaves your servers at all. Either way there is no vendor-side copy to request the deletion of.
Can it run with no internet access at all?: Yes, as an on-premise deployment. The model is served on your own servers, by Ollama or another self-hosted runtime, alongside everything else, so no egress is required at all. We configure it with you during deployment. It is not theoretical: the platform has run fully self-hosted, generation, embeddings and reranking included, and on AWS Bedrock, without losing a feature. The model configurations.

Access & security

How is passage-level access control different from folder permissions?: Folder permissions make a document either readable or not. Anything sensitive then has to be split into a second file, kept in a second place, and kept in sync by hand. Provbl attaches rules to the passages themselves, so one document serves everyone and each reader sees only the parts their group is entitled to.
Could someone get restricted content by rephrasing the question?: No. Restricted passages are excluded inside the search itself, not filtered out of a finished answer. They are never read into the response, so there is nothing to surface by asking differently.
Does it work with our single sign-on?: Yes. Authentication runs through your existing identity provider, so there is no second directory to maintain and no orphan accounts. Custom two-factor protection sits alongside it, and you decide which groups must use it and how often it is re-checked.
Who can see what someone asked?: Usage and consumption are visible to the administrators you designate, inside your own deployment. Nothing is reported back to us.

Architecture, retrieval & accuracy

Where exactly are permissions enforced?: Inside the search itself, before anything reaches the model. Every indexed passage carries the IDs of the groups allowed to read it, every query carries the user's groups, and the filter runs on both the semantic and the lexical search before their results are merged, so the model never sees a passage the user can't read. A second check after retrieval drops any mismatch and logs it as an incident, deliberately: a control that lives in one place is one refactor away from being bypassed. A user with no groups never reaches the database, and tools such as spreadsheet calculation and exact search can only touch sources retrieval has already authorised. How permissions are enforced.
What happens when someone is removed from a group?: Their very next question already respects it. We don't store who belongs to which group: your system sends the user's groups with every request, so there is nothing to sync and nothing to go stale. How permissions are enforced.
And when a document moves to another group, or is deleted?: Your source system calls a dedicated endpoint that updates only the access tags in the index, so the change applies in seconds, with no re-parsing and no re-embedding. Deleting a document removes every one of its passages. One limit, stated plainly: this relies on that event being sent, and there is no reconciliation job yet to catch drift if an event is lost. It is on the roadmap. How permissions are enforced.
How modular is it? Can we swap the model?: The generation model sits behind a neutral internal contract; Bedrock and Ollama are adapters, so the tools never see the provider and the model can be swapped without reindexing. No lock-in to one model host. Embeddings and reranking are swappable by editing a module, but changing the embedding model means re-indexing the corpus. The tools (spreadsheet formulas, document generation, charts, web page reading) are ours, described in standard JSON Schema and independent of any model; the vector store is where we are most coupled, using Qdrant's hybrid search and payload filtering directly, so replacing it is a real refactor. The platform has already run fully self-hosted and on Bedrock without losing a feature. The architecture, layer by layer.
Is it only semantic search?: No. Every search is hybrid: dense semantic search, which works across languages, plus BM25 lexical search, fused by reciprocal rank fusion. The two halves search different text. The semantic half gets the conceptual phrasing and the lexical half gets the same query with exact identifiers added, so error codes and part numbers are matched literally, not “understood”. Searches can also be narrowed by group, source type, document and page range. Retrieval in depth.
What happens with a compound question, or one the documents can’t answer?: Each message becomes a typed plan: intent, language, each separate factual requirement and any format constraints. Every requirement gets its own search and its own reranking by a dedicated cross-encoder with a relevance threshold, so when nothing passes, it says the answer isn't in your documents rather than handing you the closest wrong passage. A coverage ledger marks each requirement as covered, missing from the documentation or needing clarification, and the answer has to declare the gaps. Retrieval in depth.
How do you measure answer quality?: Retrieval is measured today: recall and MRR against a hand-labelled set, with thresholds that block any retrieval change that regresses. Every stage is timed, from planning to generation, along with time-to-first-token, tokens per second, p50 to p99 latency under load and cost per request. Automated scoring of groundedness, answer relevance and hallucination rate does not exist yet; the evaluation set is small, and that is the next priority. Until then the defences are architectural: the relevance threshold, the coverage ledger, deduplication so one passage can't pose as two confirming sources, and an optional verifier that checks the answer sentence by sentence and removes unsupported claims. We observe far fewer hallucinations than with the same model used directly, and we call that an observation, not a measurement. Quality, in full.
How do citations work?: Every claim links to the document, section and page, plus the exact fragment that supports it, not the whole passage around it. Pages are real provenance from parsing, scanned documents read by OCR included. In generated PDF and Word files the references become a Sources section, and literal search tools return exact offsets in the indexed text when you need to know precisely where something is said. Quality, in full.

Documents & sources

What file types can it read?: Documents, contracts, procedures, reports and spreadsheets, including scans and photographed pages. Advanced OCR reads headings, tables, columns, stamps, signatures and handwriting, rather than flattening a page into a wall of characters.
Which systems can it connect to?: Google Drive, email and shared mailboxes, Microsoft OneDrive and SharePoint, and databases and internal APIs. CRM, ERP and other internal tools are connected through scoped integration projects: you choose which, we build the connector.
Do we have to migrate or reorganise anything first?: No. It sits next to what you already have, and your current systems keep running exactly as they do today. The library does not need cleaning up before you start. Bulk tagging and the built-in document manager are there to sort it out afterwards.
How do we stop it answering from an outdated document?: The built-in document manager handles versions. When a new version is published the old one is retired, and a retired version drops out of answers, so nobody gets last year's procedure quoted back at them with a confident citation.

Models, cost & operations

Which AI model does it use?: The one you choose, running wherever you choose, as long as that infrastructure supports it: AWS Bedrock in your own account, Ollama on your own servers, or any other infrastructure that can serve the model. Your content is never used for training; on Bedrock it is not retained, and with a self-hosted model nothing leaves your servers. The newest frontier models are off by default, because some of them require the provider to retain prompts and outputs for up to 30 days. The generation model can be switched later without reindexing; changing the embedding model does mean re-indexing the corpus. The model configurations in full.
Do the newest frontier models really retain our prompts?: Some of them do, for up to 30 days for safety review, and that applies even inside your own infrastructure. It is a condition the model provider sets, not something the platform can switch off. That is exactly why the option is off by default and enabled only with your written sign-off. If you want the latest capability, you make the trade knowingly; if you do not, the default configuration retains nothing.
What do we get for our DPO's records?: A one-page data-flow record, shipped with every deployment, naming the exact model, the region and the retention setting in force. It is written to drop straight into a register of processing activities rather than needing to be reconstructed from a contract.
Is this a subscription? What happens if we stop paying?: It is a perpetual licence with the source code delivered, not a seat in a service we operate. The install is yours and keeps running. Our team still sets it up and supports it, but owning the licence means a change to our commercial terms cannot change your compliance posture.
Our staff already use ChatGPT on their phones. Does this help?: That is the usual reason organisations call us. 45% of employees now use AI regularly on work devices, and two in three of them go through accounts their employer does not control (Verizon 2026 Data Breach Investigations Report, pp. 12 and 60). A ban does not change that; a sanctioned tool with permissions, an audit log and SSO does, because it is better at the job than the thing they were using unofficially.
How do we keep the cost predictable?: Every account and group carries its own token budget. You see live consumption per person and per team, set caps, and choose what happens when one is reached: pause, warn, or let an administrator raise it. It also gives you chargeback figures per department.
How long does it take to get running?: Under five days to be live in your own infrastructure. On your own servers, we plan the timeline and configure the setup with you during deployment. Connecting sources and defining passage rules then happens at your pace, team by team.
Who operates it once it's live?: We do. We set it up and run it for you, and nobody on your side needs to learn anything new.

Regulation

How does this help with GDPR?: GDPR fit here is architectural. The index, the document text and all processing run inside your own deployment, and the model runs in your own infrastructure too, on AWS Bedrock in your own account, Ollama on your own servers, or wherever else you choose to serve it. Your content is never used for training. No third-party AI vendor ever sees the data. For special-category data such as health records, that difference is what makes internal AI possible at all.
What about the EU AI Act?: Cited, traceable answers and controlled data access make internal AI use straightforward to document as the Act's rules take effect. Every answer already carries the document, page and passage it was built from.
Our clients contractually forbid putting their data into AI. Does that rule this out?: That clause is usually about public AI tools and third-party AI vendors. Here their data never reaches one: it stays under your access rules, in your own infrastructure and EU region, and the model that reads it runs in that same infrastructure. That is what typically lets you adopt AI without renegotiating the contract.
Is this a medical device?: No. In healthcare settings it helps people find and verify information in your own documents. It does not make medical decisions and it is not a medical device.

Still have a question? Ask it live

Book a demo

30-minute demo, on your files, no commitment.