Rootnia

Most business information lives in PDFs, scans and images a program cannot read directly. Extraction turns those into structured fields: it runs OCR where needed, then uses a schema to pull the specific values you care about — invoice total, parties, dates — and validates them against rules. It is guided step by step so a person stays in control when a document is ambiguous.

Instead of a single "chat with your PDF" box, each document gets a session where you switch task: ask a question, reason over the content, analyze or classify it, extract fields, compare two versions, assess risk, or preview an anonymised copy. Every session is scoped to one tenant, so one customer’s documents never bleed into another’s.

Before you extract, you often need to know what a document IS — an invoice, a contract, an ID. A cascade runs cheap, certain methods first (patterns, then retrieval, then a small model) and only escalates to an LLM arbiter when needed, which is faster and more auditable than asking a big model every time. Honest scope: today it is a standalone feature, not yet wired into the extraction step.

Real document work is not one upload at a time — it is thousands of files arriving in a bucket or a SharePoint. Extraction can run as a scheduled, queued job over connected sources, reusing the workflow engine. Crucially, a source with no credentials reports REQUIRES_CONNECTION instead of pretending to have run, so you never get fabricated results.

RAG lets the assistant answer from YOUR documents rather than its training data, which cuts hallucination and stale knowledge. "Hybrid" means it combines semantic (vector) search, which finds related meaning, with keyword search, which catches exact terms; a reranker then orders the best passages. "Multi-tenant" means each customer’s knowledge base stays isolated from every other.

A number is only trustworthy if you can see where it came from. An anchor links each extracted value to its exact place on the page, so a reviewer can click a figure and jump to the source. Honest scope: the anchor is set by the user (drag-and-drop), not auto-derived from OCR coordinates — we do not over-claim automatic grounding.

Lineage answers "how did this record come to exist?". Every ingestion is recorded with a provenance envelope, and the system builds a graph you can query to trace a value back through its transformations. This is distinct from observed data lineage (OpenLineage-style, across pipelines), which is still planned.

An LLM can state something false with confidence, so the verifier jury checks it. For every extracted or generated value it compares the claim against the source and returns a verdict — Supported, Invented, Review or Unverified — using deterministic rules with NO model in the verification path, so the checker cannot hallucinate the way the generator can. It is the anti-hallucination backbone and the most heavily tested part of the platform.

A useful assistant does more than talk: it decides which tool to call — search the docs, run a query, start a workflow — and with what arguments. The router picks the right tool, sensitive actions wait for human approval, and every call respects the user’s role. "Hybrid" means it mixes grounded retrieval with tool actions in one conversation.

This turns a plain question — "revenue by month last year" — into a database query. To make it reliable it is given the real schema, guided by a bank of verified example question→SQL pairs, and restricted to read-only SELECT. The generated SQL is shown to you so the answer is auditable, and if the query returns no rows it says so instead of inventing numbers.

When a query returns a table, the system picks a sensible chart — bar, line, pie — automatically, so the answer is readable at a glance rather than a wall of rows. It is a presentation layer over the same verified query result.

The outputs — a report, an extracted dataset, a summary — export to the formats people actually work in: Word, PDF, Excel and PowerPoint, so results leave the platform in a usable, shareable shape.

For hands-free or accessibility use, speech-to-text transcribes what you say into the chat and a document-to-audio path reads a document back. It is an input/output channel over the same answer engine, not a separate assistant.

An "agent" here is not any call to an LLM: it is a component with a defined job, a stable output contract and verifier-jury grounding so its answers are checked. The audit counts 28 real agents; the larger "93" seen elsewhere is the palette of workflow NODES in the builder — a different thing — stated plainly so the capability is not inflated.

A workflow is modelled as a directed acyclic graph: nodes are tasks, edges are dependencies, and because there are no cycles the run has a clear order and is guaranteed to terminate. The engine runs independent branches in parallel, retries failed nodes, and can pause at an approval gate for a human — real orchestration, not a linear script.

For larger jobs several agents cooperate, but without governance that invites runaway behaviour. A coordinator assigns work under an execution budget (so it cannot loop forever), routes outputs through the verifier jury, and holds sensitive steps for approval — collaboration with brakes.

Standing up a vertical — say accounting or legal — from scratch is slow. A pack installs a curated bundle (the relevant agents, workflows and connections) in one step. It is a template installer only: no billing and no third-party services are activated by installing a pack.

When agents can take real actions, a human needs one place to say yes or no. The inbox unifies everything waiting on approval — workflow steps, agent runs, agent actions, chatbot tool calls — and enforces separation of duties: whoever proposed an action cannot be the one who approves it.

A2A lets agents discover each other and delegate tasks; MCP standardises how an agent reaches tools and context through a common registry. The transport (HTTP/SSE/OAuth/federation) exists and is tested in-process; wiring it for production — mutual TLS, a real identity provider — is still Planned, stated so the maturity is not overstated.

To put the assistant on your own site or app there is a chat widget and a server-to-server API. Authentication is by API key, and the caller’s identity is taken from the key itself, never from a field in the request body — which closes the classic impersonation gap (IDOR). Keys are managed in-platform.

An inbound WhatsApp webhook lets users reach the same answer engine from a channel they already use; incoming requests are signature-verified. It is an honest scaffold — it works once real credentials are loaded, and without them it does not pretend to be live. (No Telegram integration exists.)

Migrating code from a legacy language is only safe if the new code behaves the same. Differential testing runs the original and the migrated version on the same inputs and compares outputs; a mismatch is a bug, which an agentic swarm then repairs and re-tests. "Verified" means there is a trace that behaviour matches — not a hopeful translation.

Moving data between systems is where silent corruption hides. A data-engineering engine runs quality rules — completeness, uniqueness, validity, consistency, freshness — so problems are caught and measured rather than discovered later downstream.

Adapters to big-data engines broaden where Rootnia can run heavy workloads. Honest scope: several of these "breadth" adapters are connection-tested only — they prove the link works — but are not yet operated end-to-end, so they are marked experimental, not live.

Three layers of access control: two-factor authentication to prove who you are, role-based permissions (a hierarchy of what each role may do), and tenant isolation so one customer’s data and actions never reach another’s. The identity used for every check comes from the signed token, not a field in the request — which prevents the IDOR class of bugs — and cross-tenant isolation is certified by tests.

Agentic systems add attack surface. A default-on prompt-injection guard screens untrusted input before it reaches an agent, and an SSRF guard checks every outbound call (deny-by-default to internal addresses, anti DNS-rebinding). Honestly: the injection guard is heuristic and declared evadable — defence in depth, not a silver bullet — which is why authority and approval also sit outside the model.

Per tenant, the dashboard reports value generated, hours saved, AI cost and net ROI, computed from audit, billing and execution data. The savings figure is an explicit, configurable estimate — it assumes, for example, minutes saved per document — not an empirical measurement, and the page says so, so it is not mistaken for a measured result.

Governance needs a record that cannot be quietly edited. Every consequential action is written to an append-only log chained by hashes, so who did what, when and over which data is reconstructible and tamper-evident.

Before documents are shared or used for analysis, personal data can be detected and redacted. Detectors cover Argentine and EU personal-data patterns over a redaction layer, so sensitive fields can be removed or masked.

For a multi-tenant product, each tenant has usage quotas and limits so one cannot starve the others, plus the data-subject rights and retention controls that GDPR requires.

Standards-based digital signatures — PAdES for PDF, PKCS#7 — would let documents carry a legally recognised signature. This is designed but not yet implemented, marked Planned rather than implied to exist.