The vocabulary, in plain English

Intelligent document processing is the use of AI to read, understand and extract meaning from documents — including scans, PDFs and unstructured files — going beyond OCR's raw text capture to interpret context, entities and relationships.

RAG is an AI architecture where a language model's answers are grounded in documents retrieved from your own corpus at question time, rather than relying on what the model memorised in training. It's how systems like Marella answer from your material instead of the open internet.

Private RAG is retrieval-augmented generation run under your organisation's control: your documents stay in an isolated environment, are never used to train shared models, and access follows your permissions. The alternative — public AI tools — offers no such guarantees.

AI citations are references attached to an AI-generated answer that point to the exact source passages the answer relies on, so a human can verify it in one click. Citations turn 'trust me' into 'check me'.

Grounding means constraining an AI system to answer only from verified source material. A grounded answer can show its evidence; an ungrounded one is the model's best guess.

A hallucination is a confident but false statement produced by an AI model. Grounding plus citations is the practical defence: if the documents don't support an answer, the system should say so.

A knowledge network is a linked map of the entities and concepts inside an organisation's documents — people, companies, topics, ideas — maintained automatically and curated by the organisation. Unlike a search index, it accumulates structure and improves with use.

An ontology is the set of concepts and categories an organisation uses to describe its world, and the relationships between them. In Marella, the ontology is curated by your team, so the AI reasons with your vocabulary rather than a generic one.

Entity extraction is the automatic identification of the people, organisations, places, documents and topics mentioned in text. It's the raw material from which a knowledge network is built.

Enterprise search is the retrieval of information across an organisation's internal systems. Modern enterprise search is semantic (meaning-based) rather than keyword-only, and increasingly answers questions directly with citations instead of returning ten links.

Semantic search retrieves content by meaning rather than exact keywords — a query about 'notice periods' finds clauses about 'termination windows'. It's typically powered by vector embeddings of text.

AI governance is the set of controls that make AI use accountable inside an organisation: who may use it, on which data, with what oversight, and how activity is recorded. Practically it means access control, scoping and audit trails — built into the platform, not added by policy document.

RBAC grants each person capabilities according to their role, rather than individually. In document AI, it decides who can ask questions of which document sets — mirroring the permissions your filing system already enforces.

An audit trail is a tamper-evident record of who did what, when — in an AI platform: who asked which questions, which sources were accessed and what was answered. It converts 'who saw this?' from an investigation into a query.

Data sovereignty is the principle that data is subject to the laws and controls of the place and organisation it belongs to. For UK organisations it typically means UK hosting, UK GDPR compliance and the option to run systems inside their own environment.

A private deployment runs the AI platform inside an organisation's own environment rather than a vendor's shared cloud — the strictest answer to data-residency and sovereignty requirements.

ISO 27001 is the international standard for information security management systems (ISMS). Certification means an accredited auditor has verified an organisation's security controls and processes — a common procurement requirement in regulated sectors.

UK GDPR is the United Kingdom's data protection regime, enforced by the ICO. For AI systems it governs the lawful processing of personal data in documents — making scoping, minimisation and access control platform requirements, not options.