Classification
A customer complaint becomes a category from the list YOU defined. A label outside the list is rejected — the model cannot invent a new one.
Intelligence layerThe smart part.
The smart part.
And why you can trust it.
Seven mechanisms wired into one layer. Each works on its own, and each is built so the model is never the source of truth.
Why the model can be trusted
The model never writes to the database directly — your schema stands between them.
AI agents inside your processes
Called from a trigger, a card button, a schedule or a bot.
Schema-checked extraction
Text or a photo becomes an object with the fields you declared. Types are coerced, extras dropped, the schema is yours.
Prompts as configuration
A new skill is a row in settings, not a function in code. Wording changes without a release, with a “try it” panel and a token counter.
Honest refusals
“Provider not configured”, “quota spent”, “this model has no vision” — three distinct reasons, not one vague “AI unavailable”.
Usage and budget
Every call records tokens under the provider that actually answered. A monthly cap stops spending in a way you can read.
Your choice of model
ollama, Anthropic, DeepSeek, OpenAI — with a per-call override: a cheap model to classify, an expensive one to read documents.
Telegram bots that read documents
The shortest path from a photo to a record.
Receipt photo → expense
The bot takes a photo, extracts line items and totals, and creates the record through the same path as manual entry: validation, workflow, audit.
The person behind the chat
The bot knows who is writing and acts under THAT person's permissions — it never sees more than they do.
Long work in the background
Recognition can take two minutes; a bot has fifteen seconds. A queue accepts the job and runs it on the full budget.
Confirmation before writing
The bot shows a card with what it read and waits for a button. Nothing reaches the database silently.
Knowledge base and meaning-based search
A record becomes a sentence — and starts being found.
Record → connected prose
A card rule assembles a readable description from fields and relations. That is what search sees, not the raw table.
Hybrid search
Words and meaning together: a lexical index and vectors fused into one result list. Without an embedding provider it matches words — and says so.
Answers with citations
“Ask AI” answers with links to your own records. Permissions apply BEFORE the prompt is built, so a citation cannot leak what you may not see.
Similar records
“Has this happened before?” — from the document's rarest terms and its own vector, with no call to a model at all.
Public help centre
The same content as an open help site on your domain, with private spaces behind a share link.
Auto-drafting
Summaries, tags, “see also” and article drafts — the model proposes, a human accepts item by item.
Ask your data in plain words
The knowledge base answers “what”. This answers “how much”.
Structure, not SQL
The model writes no database queries. It fills in a config the system validates against YOUR schema, and the existing report engines do the counting.
The explanation next to the number
“Sum of «Amount» in «Operations» for 01.07–31.07” — built from the very config that ran, so the two cannot diverge.
Asking back instead of guessing
If the model hesitates between two entities it asks with buttons and executes NOTHING.
Three layers of permissions
Rows, entity and fields. A hidden field is never summed, and a refusal does not enumerate someone else's schema.
Semantic links and inference
Statements about two records — and what follows from them.
What no form can hold
“Caused”, “duplicates”, “supersedes”. The predicate registry is yours: name, reading in both directions, allowed types.
Inference on demand
Transitive chains of causes, duplicate clusters, “advances with no report” — as datalog rules, not another table.
A model's guess goes in a queue
What the AI suspects is filed as a proposal and, until a human confirms it, exists neither for inference nor for search.
MCP: an agent builds the platform
Not “generate code” but “do it in the system” — under your permissions.
Over 140 tools
Entities, fields, reports, dashboards, bots, the knowledge base, the site, document templates — anything done by mouse can be done by dialogue.
The agent's rights are your rights
The agent acts as its token's actor. A hidden feature does not exist for it: the tool simply is not in the list.
One principle across the whole layer
The model is never the source of truth. It either turns text into structure the system then validates, or answers with facts it did not retrieve itself.
The consequences show everywhere: an invented field fails validation, a label outside the list is rejected, a citation cannot point at a record you cannot see, and a link proposed by AI waits for confirmation.
And conversely — where a model is not needed, there is none. Similar records, rare-term ranking, language normalisation and datalog all run without a single provider call.
Frequently asked
Which AI providers are supported?
ollama (local, offline), Anthropic, DeepSeek and OpenAI. Provider and model can be overridden per call — the rest of the system does not notice.
Can it work without any cloud AI?
Yes — a local ollama gives both generation and embeddings, so data never leaves your server. Meaning-based search works the same way in that mode.
What happens when the model gets it wrong?
Output is checked against a set: a label outside the list, a field outside the schema or a non-existent entity are rejected before execution. The model never writes to the database directly.
See it on your own data
The smart part switches on separately from the rest — and off just as easily.
Questions, feedback or need a hand — gloud@gloud.com.ua