The product
Eight areas in the left-hand menu, three of them shown here in full: what your team types, what comes back, where the files live and who is allowed to reach them. The figures are not placeholders. Every one of them is from an install that is in use.
Answer
I can't draw a detailed comparison between July and June from the accessible files: they contain attendance data only, not the context needed for a meaningful comparison.
What the attendance data does show [1][2][3][4][5][6]:
Structure:
Limitations. The files contain nothing on workload, productivity, specific projects, deliverables or other performance metrics.
To compare the two months properly you'd need activity reports or performance data, which aren't in the current files. Which aspect would you like me to compare?
Sources
Copy 6.3 s · 11,591 tokens
Ask a follow-up…
Today
Payroll statements for July
Compare attendance, June vs July
What's in the supplier contract?
Which invoices came in July?
2 days ago
Onboarding brief for the new hire
A real answer, including the part most demos hide: the model saying the files can't support the question, and naming what's missing.
Chat
Files
Upload documents and scans for AI search, then control who can access them.
Folders
All files 10 items
| Name | Access | Author | Size |
|---|---|---|---|
| Engineering | Engineering | admin | 32 items |
| Finance | Finance | admin | 27 items |
| HR | HR | admin | 31 items |
| IT Security | Programming IT Security | admin | 0 items |
| Legal | Legal | admin | 30 items |
| Management | Management | admin | 32 items |
| Marketing | Marketing | admin | 30 items |
| Programming | Programming | admin | 0 items |
| Sales | Marketing Finance | admin | 31 items |
| Support | Support | admin | 30 items |
Knowledge
Access groups
Access groups
Organize users into reusable groups, then grant those groups access to files.
| Group | Users | Files |
|---|---|---|
| No access Default No access to files or workspace content. | No users | 0 |
| Full access Default Full access to every file and workspace feature. | 2 · 2 can edit | 281 |
| Engineering Engineering files | 1 · 1 can edit | 32 |
| Finance Files for the fictional test business | No users | 96 |
| HR hr files for fictional business | No users | 31 |
| IT Security security files | 1 · 1 can edit | 0 |
| Legal Legal files | No users | 30 |
| Management Management files | No users | 32 |
| Marketing Marketing files | No users | 61 |
| Programming | No users | 0 |
| Support external support files | No users | 30 |
Straight answers
Running it in your own infrastructure, not a vendor's, is what makes it permissible. The nine questions below are what makes it useful, including the one about what happens when it does not know.
PDFs, Word documents, spreadsheets, and scans or photographed pages that were never digital text to begin with. Spreadsheets keep their rows and columns, so a figure stays attached to the column it belonged to. That is what lets a citation land on the right cell rather than somewhere in the right file.
It reads spreadsheets and can compare, aggregate and summarise what is in them, and it cites 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 human confirms it.
Only when you switch it on. There is an Internet toggle on every question, off by default. Your documents never go out through it. It only lets an answer reach outside when you decide a question warrants it.
It says so, and says why. The chat screen at the top of this page shows the product declining to compare two months, because those files hold attendance codes and nothing about workload or output. It then lists exactly what data would be needed. A tool that answers everything is a tool you end up double-checking every time.
The one 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.
Yes. Integrations and Plugins are built into the app. Google Drive, email and shared mailboxes, OneDrive and SharePoint, databases and internal APIs are standard. CRM, ERP and in-house tools connect through plugins, built as scoped projects: you pick which, we build it, read-only by default.
Four admin areas in the same app: Users (accounts and roles), Access groups (who reaches 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 a number you look up rather than one that arrives.
Yes, and it exports them. Ask for a comparison, a roll-up of findings or a briefing and you get a finished document out as PDF, Word or Excel, assembled from your own material rather than written from a blank page. Ask for a chart and it draws one from the figures it just cited; the visual travels with the file.
You decide, down to the passage. Groups carry file access; users carry groups. Two people can ask an identical question and get different answers, and the passages one of them isn't entitled to are never retrieved at all. How access is enforced.
How it works
Parsing, OCR, the index, lexical search and the access rules run as containers on your own deployment. The one step that calls a model goes to the model running in your own infrastructure: AWS Bedrock in your own account, Ollama on your servers, or wherever else you choose to serve it. No third-party AI vendor sees your documents, and nothing that comes out is unattributed.
Files are parsed, split into passages and indexed on your own deployment, with page and passage positions kept. Every passage is stored with the IDs of the groups allowed to read it.
Every question carries the asker's groups, and the filter runs inside the search itself, before anything reaches the model. A second check after retrieval drops any mismatch and logs it. How it is enforced.
A model running where you choose: AWS Bedrock in your own account, Ollama on your own servers, or any other infrastructure that serves it. Your content is never used for training. Wherever it runs, it answers only from passages the asker may see, and says so when the answer is not in your documents.
Storage, search and the model call all stay inside this line
Never used for model training · never sent to a third-party AI vendor
How it gets used
Day one · the administrator
Upload. Drag in the folders you already have, including scans that were never searchable. The install shown here holds 281 files.
Create groups. Engineering, Finance, HR, Legal, IT Security, Support. Two exist already: No access and Full access.
Attach files to groups. A folder can answer to more than one group. Sales files here sit under both Marketing and Finance.
Add people. Assign each account its groups. Tudor gets Engineering and IT Security; he will never see the other 249 files.
Every day after · everybody else
Type the question the way you'd ask a colleague. Narrow it to specific groups if you want, or leave it on All groups.
Read the answer with numbered citations inline, and the source files listed underneath with the count of passages used.
Open the source if it matters. Expand a file and you get the cited passages themselves, not a page reference to go hunting with.
Keep going. Follow-ups stay in the thread, and the thread stays in your history, searchable days later.
No third-party AI vendor
The model runs in your own infrastructure, on Bedrock, Ollama or another host you choose. Nothing is forwarded to a consumer AI tool.
No training on your content
On Bedrock the passages are not retained or trained on; on your own servers they never leave them. Your documents do not become training data, for us or anyone else.
No document store elsewhere
The index, all document text and the originals stay on your deployment. The model receives only the passages an answer needs.
No uncited claim
If it can't be traced to a passage, it doesn't get presented as fact.
Where it installs and the three rollout steps are on Deployment. How the boundary holds up under audit is on Security.
Retrieval
Any model can write fluent prose. What decides whether it is right is what it was handed to read. So before a model sees anything, each question is planned, searched two ways, filtered, reranked and checked for gaps.
Illustrative example · one question, three requirements
The question
“What does alarm E-217 mean on the chiller, how do I reset it, and who approved the last firmware change?”
Requirement 01
What alarm E-217 means
Search · semantic phrasing, plus “E-217” matched literally
Rerank · 3 passages clear the threshold
Covered
Requirement 02
How to reset it
Search · reset procedure, same filters
Rerank · two procedures, for two chiller models
Needs clarification
Requirement 03
Who approved the firmware change
Search · approvals, change records
Rerank · nothing clears the threshold
Missing from your documents
The answer explains E-217 with citations, asks which chiller model before giving a reset procedure, and states plainly that no document records the firmware approval.
Dense semantic search, multilingual, runs beside BM25 lexical search, and the two are fused with reciprocal rank fusion. The semantic half gets the conceptual phrasing; the lexical half gets the same query with the exact identifiers added. An error code or a part number is not “understood”. It has to match literally.
Each message becomes a typed plan: intent, language, the separate factual requirements and any format constraints. Every requirement gets its own search and its own reranking, so a compound question gets every part answered, not just the first.
Searches narrow by access group, source type, document and page range. The group filter runs inside each search, before results are fused, so the model never sees a passage the user isn't allowed to read. How access is enforced.
A dedicated cross-encoder rescores the candidates for each requirement. If nothing clears the threshold, you are told that it isn't in your documents, instead of being handed the closest wrong passage.
Every requirement is marked covered, missing from the documentation or needs clarification, and the model is instructed to declare the gaps in its answer. A missing part is named, not quietly skipped.
Accuracy & citations
A citation you can open is worth more than a promise of accuracy. So the product shows its sources down to the fragment, is built so that gaps get declared rather than filled, and we tell you exactly which parts of its quality we measure today.
A cited claim
Alarm E-217 is a condenser high-pressure trip, raised when discharge pressure exceeds the set limit.
Chiller_Service_Manual.pdf · § 6.3 Alarm codes · p. 48 · scanned, OCR
“…E-217: condenser high-pressure trip. Unit stops; manual reset required…”
Illustrative example.
A relevance threshold
It prefers “that isn't in your documents” to building an answer on a weak passage.
A coverage ledger
Gaps are declared per requirement, so a missing part can't be papered over.
Deduplication
One passage can't appear as two sources confirming each other.
An optional verifier
Before delivery, it audits the answer sentence by sentence and removes what isn't supported.
Measurement
In use, we see far fewer hallucinations than from the same model used on its own. That is an observation, not a measurement, so it carries no number here.
Measured today
Retrieval quality, as a gate
Recall and MRR against a hand-labelled set. Thresholds block any retrieval change that makes them worse.
Latency, stage by stage
Time spent in planning, embedding, retrieval, reranking and generation, plus time to first token and tokens per second.
Behaviour under load
Load tests with p50 to p99 percentiles, and the cost of each request.
Not measured yet
Automated answer scoring
No automated scoring yet for groundedness, answer relevance or hallucination rate.
A larger evaluation set
The labelled set is small today, and we say so.
Next on the roadmap
Both are the next priority. Until they exist, we don't quote an accuracy figure.
Architecture
The platform has run entirely on self-hosted models, with generation, embeddings and reranking all local, and on AWS Bedrock, without losing a feature. Here is how modular each layer is, including the one that isn't.
Generation model
Fully abstracted behind a neutral internal contract. Bedrock and Ollama are adapters; the tools never see the provider.
Swap the adapter (3 of 3)
No reindexing
Tools
Excel formulas, document generation, charts and web page reading are ours, described in standard JSON Schema.
Model-independent (3 of 3)
Nothing to change
Embeddings & reranking
Swappable by editing a module, not through a plug-in.
Edit a module (2 of 3)
A new embedding model means re-indexing the corpus
Vector store
Qdrant, used directly for hybrid search, RRF fusion and payload filtering. The most coupled layer.
A real refactor (1 of 3)
Not a configuration change
Your cloud: AWS Bedrock, in your own account
Inference runs in your own account and region. Bedrock does not retain what it is sent or train on it, and no third-party AI vendor sees your documents.
Your servers, or anywhere else: Ollama and other hosts
Run the model on your own servers with Ollama, or on any other infrastructure that can serve the model you pick, and nothing leaves at all. We configure it with you during deployment.
Features
Reading documents properly, controlling what each account spends, deciding who sees what, proving who is asking, pulling from every system you already run, keeping the library current, and choosing the model yourself. In the app, most of them live under Knowledge, Access groups, Configurations, Usage, Integrations and Plugins.
01 · Advanced OCR
Most of what an organisation knows is locked in scans, photographed pages, signed PDFs and spreadsheets exported to print. Plain text extraction gives you a wall of characters with no idea which number belonged to which column.
Provbl reads headings, tables, columns, stamps, signatures and handwriting rather than bare characters, so a passage keeps the meaning it had on the page. That is what makes a citation land on the right paragraph instead of somewhere on the right document.
Reading a scanned page
§ 4.2 TERMINATION
notice_period = 60 days | framework = 90 days
signed_by = "M. Ionescu" (handwritten)
Token budgets · this month
02 · Cost control
Internal AI usually fails its second budget review, not its first. One team discovers the tool, starts running everything through it, and the invoice arrives with no way to tell who spent what, and no way to stop it happening again.
Every account and every group carries its own token budget. You see consumption per person and per team, set caps, and decide what happens when one is reached: pause, warn, or let an administrator raise it, so the invoice is something you look up rather than something that arrives.
03 · Permissions
Tag once, control everywhere. Documents and passages carry tags such as commercial, patient data, board only or site 4, and access rules are written against those tags and your user groups, not against folder trees that drift the moment somebody reorganises a drive.
Because the rule sits on the passage, the same document serves everyone: each reader gets the parts their group is entitled to, and the restricted parts are never retrieved at all.
Tags applied · rules enforced
Connecting your sources
04 · Integrations
Knowledge is never in one place. It's in Drive, in OneDrive and SharePoint, in the shared mailbox nobody has cleaned out since 2019, and in the database behind an internal tool. Provbl connects to them where they are.
Connected sources stay in sync, so the index reflects what your teams work from day to day. Every connector runs on your own deployment, under the same access rules as everything else.
05 · Custom connectors
The systems that matter most are usually the ones nobody else integrates with: the ERP that was configured for you in 2014, the CRM with fourteen custom objects, the internal tool one person maintains.
You choose which of them to connect, and we build the connector as a scoped project. It is specified and agreed before any work starts, with the access boundary written in. No migration, no replacement: your systems keep running exactly as they do today.
Scoped connector · built for your system
Model configuration
Answer length & retrieval depth
06 · Model configuration
The model sits behind a neutral internal contract, so it is a choice, not a lock-in. Run it on AWS Bedrock in your own account, on Ollama on your own servers, or on any other infrastructure that can serve the model you pick. Your content is never used for training, and no third-party AI vendor sees your documents. Different departments can run different models against the same index.
The newest frontier models are available too, but some of them require the provider to retain prompts and outputs for up to 30 days for safety review. That configuration is off by default and enabled only with your written sign-off, and whichever you choose is written into a one-page data-flow record for your DPO.
Where the model runs is your call, and we configure it with you during deployment. It is not theoretical: the platform has run fully self-hosted and on Bedrock without losing a feature. Retrieval depth, answer length and behaviour are yours to tune. When a better generation model ships, you switch to it without reindexing; changing the embedding model does re-index the corpus. What swaps easily, and what doesn't.
07 · Access security
Passage-level rules only mean something if the person behind the account is who they claim to be. Provbl adds a second factor at sign-in, configured the way your organisation works rather than the way a vendor decided.
Choose which factor to use, which groups must use it, and how often it is re-checked. You can make it stricter for the roles that reach commercial, clinical or board-level passages and lighter for everyone else. It sits alongside your single sign-on rather than in front of it.
Signing in · two factors
Document manager · bulk actions
08 · Knowledge management
An AI assistant is only as current as the library behind it. Provbl ships with a full document manager, so the library is maintained inside the same app that answers from it, rather than in a spreadsheet somebody keeps alongside.
Upload, organise, tag in bulk, publish a new version and retire the old one. A retired version drops out of answers the moment it is superseded, so nobody gets last year's procedure quoted back at them with a confident citation.
Capabilities
54 capabilities · 6 groups
None of these change when the industry does. A hospital and a contractor get the same platform; what differs is which documents go into it and which words come back out. See it across the industries.
Use cases
The same question box that returns a one-line answer will also compare two contracts, pull a year of findings into one list, or produce a document you can send out. Every output carries the documents and pages it was built from.
Ask
The everyday case: someone needs a number, a clause or a rule, and needs to know it's the version in force today rather than the copy from two years ago.
"What notice period applies to our supplier contracts?"
60 days' written notice, per Section 4.2 of the Master Services Agreement, except framework suppliers (90 days).
Supplier_MSA_2024.pdf · p. 12 · v7 · current
Compare
Put two agreements next to each other and get back the differences that matter: terms, notice periods, liability caps, rebates. Ask for a chart of the numbers and it draws that too.
| Term | Master services | Framework |
|---|---|---|
| Notice period | 60 days | 90 days |
| Rebate | 2% above €500k | None |
| Auto-renewal | 12 months | 24 months |
2 documents · 6 passages cited
Roll up
Audit prep normally starts from an empty page and three weeks of reading. Here it starts from a list assembled out of the reports you already have, each line pointing at its source.
Draft
Export the result as PDF, Word or Excel: a review, a summary for a regulator, a briefing for a board, assembled from your own files rather than written from scratch.
Supplier notice periods, Q3 review · 11 citations · .docx
Onboard
Most organisations have one person who knows how things really work. When they retire or leave, six months of institutional knowledge goes with them. A new starter can question the library instead, and get the current answer with its source.
Answer clients
Security questionnaires, tender responses and RFI packs are mostly re-answering things your organisation has already answered somewhere. Pull those answers out of past submissions and policies, each one traceable to the document it came from.
Real questions from a real install
“Compare attendance, June versus July.”
Two spreadsheets, six cited passages, and an honest note that the files can't support a productivity comparison.
“Which invoices came in during July?”
The kind of question that otherwise means opening a folder and a finance system and reconciling by eye.
“Tell me about the contract.”
Answered from the version in force, with the clause quoted rather than paraphrased.
“Payroll statements for July.”
Visible to the roles entitled to payroll. To everybody else, the passages are not retrieved at all.
30-minute demo, on your files, no commitment.
