Search and RAG
How a question turns into an answer with sources.
Enterprise AI glossary · 45 definitions
Enterprise AI terms in plain English.
Plain-English definitions of the terms you will meet in a sales conversation. Each one links to the questions worth asking.
The label on a product does not tell you whether it fits. Ask what it actually does.
Find the term behind the claim.
45 definitions
Browse by topic
Four places to start
Documents and evidence
How files become text, passages, fields and inspectable citations.
Privacy and control
What a privacy label actually covers.
Models and agents
What sits behind a model's answer, and where it stops.
Browse by letter
A
4 terms
Access control
Access control is the policies and technical rules that decide which people or services can reach a resource, and what each one may do with it.
Agentic AI
Agentic AI is AI set up to pursue a goal over several steps, choosing its own tools and deciding its next move inside permissions a person has set.
AI citations
AI citations connect an AI output back to its sources. A link to a file and a passage that supports the claim attached to it are not the same thing.
AI hallucination
An AI hallucination is output that sounds plausible and is unsupported or wrong. Retrieval and citations lower the risk of one without removing it.
C
2 terms
Chunking
Chunking is the splitting of source content into smaller units for indexing and retrieval. Where the boundaries fall changes what a system can find later.
Context window
A context window is how much information a model can hold in view for one request. A bigger window is not memory, and the model may not use every detail in it.
D
5 terms
Data extraction
Data extraction finds chosen values, entities or relationships in source material and turns them into structured output another system can use.
Data governance
Data governance is the rules and responsibilities that decide who owns a set of data, who may use it, how accurate it has to be and when it is deleted.
Document AI
Document AI is software that applies machine learning or generative models to documents, whether to parse them, pull out values or answer questions from them.
Document classification
Document classification sorts documents into categories using rules, a trained model or human review. Accuracy means nothing until you fix the label set.
Document parsing
Document parsing converts a file's text, layout and metadata into structures a system can search. How well it works depends on the format and the file itself.
E
3 terms
Embedding
An embedding is a numeric representation of content used to compare meaning. It makes retrieval possible without keeping every nuance or promising relevance.
Enterprise AI
Enterprise AI is AI used inside an organisation, where identity, data handling and who is accountable are requirements rather than afterthoughts.
Enterprise search
Enterprise search is search across an organisation's approved sources, and it carries identity and permission requirements that public web search never has.
F
1 term
Fine-tuning
Fine-tuning is extra training that adapts a model on chosen examples. It changes how the model behaves, which is not the same as feeding it current material.
G
1 term
Grounding
Grounding ties an AI output to chosen context or evidence. It can make an answer easier to support, but both the sources and the claims still need checking.
H
1 term
Hybrid search
Hybrid search combines more than one retrieval signal, usually keyword matching alongside semantic similarity, then merges the results into a single ranking.
I
3 terms
Inference
Inference is running a trained model on an input to produce an output, whether that output is a prediction, an extracted value or generated text.
Information retrieval
Information retrieval is the finding and ranking of material relevant to a question. Measure that separately from how good the generated answer reads.
Intelligent document processing
Intelligent document processing turns documents into structured information, combining parsing and extraction with the human review that checks the result.
K
2 terms
Knowledge graph
A knowledge graph is a structured representation of entities and the relationships between them. Marella's knowledge network adds concepts and curation on top.
Knowledge management
Knowledge management is the practices and systems an organisation uses to create, organise, share, correct and hold on to what it knows.
L
2 terms
Large language model (LLM)
A large language model is trained on very large text collections to predict and generate language. Who provides it and how it handles your data matter too.
Latency
Latency is the time between a request and an agreed response point. A useful figure says what was measured and gives percentiles, not just an average.
M
2 terms
Metadata
Metadata is information describing another resource, such as who owns it, what type it is, when it was made and how it is classified for access.
Multimodal AI
Multimodal AI takes in or produces more than one kind of content, such as text, images or audio. Which combinations work varies by model and workflow.
N
2 terms
Named entity recognition (NER)
Named entity recognition finds and labels mentions in text, such as people, organisations, locations and dates, so they can be counted or linked to records.
Natural language processing (NLP)
Natural language processing is the set of computational methods for analysing or producing human language, from classifying a document to generating text.
O
2 terms
On-premises deployment
On-premises deployment is software running on infrastructure the customer controls. Where data goes and who can support it still depend on the contract.
Optical character recognition (OCR)
Optical character recognition turns text in an image or scan into machine-readable characters. Whatever it gets wrong is carried into every later step.
P
3 terms
PDF parsing
PDF parsing pulls text, layout and metadata out of a PDF into structures software can use. Two PDFs that look alike can be encoded completely differently.
Private cloud
A private cloud is a cloud environment dedicated to one organisation. How isolated it is, who operates it and where the data sits depend on the service.
Prompt engineering
Prompt engineering is the design and testing of the instructions, context and output constraints given to a model for one particular task.
Q
1 term
Query understanding
Query understanding is the step before retrieval that works out what the user meant, then turns it into the terms and filters the system searches with.
R
3 terms
Recall and precision
Recall is the share of relevant items a system found; precision is the share of what it found that was relevant. Neither counts until you define relevant.
Reranking
Reranking is a second pass that reorders an initial set of candidate results using an extra model, rule or scoring method before any of them are shown or used.
Retrieval-augmented generation (RAG)
Retrieval-augmented generation fetches source material and hands it to a model to answer from. It grounds the answer without proving the answer correct.
S
2 terms
Semantic search
Semantic search matches on meaning rather than exact wording, so it can find the right passage when the query and the document use different words.
Structured data
Structured data is data organised to a defined schema, such as records with consistent fields, types and relationships between them.
T
3 terms
Table extraction
Table extraction turns a table into rows, columns and cells. Merged cells, headers that repeat across pages and nested layouts are where it usually breaks.
Token
A token is a unit of text or other input as a model processes it. Tokens are not words, so a limit or a cost cannot be worked out from a word count alone.
Transformer
A transformer is a neural-network architecture built around attention, and it is the design underneath most current language and multimodal models.
U
1 term
Unstructured data
Unstructured data is information with no consistent schema, such as reports, emails or transcripts. It usually still carries layout and meaning worth keeping.
V
1 term
Vector database
A vector database is built to store and search vectors such as embeddings. It is one component of a retrieval architecture rather than the whole of one.
Z
1 term
Zero-shot learning
Zero-shot learning is handling a task with no task-specific labelled examples at evaluation time. Whoever reports it should say what the model had not seen.
Three questions to ask any vendor
Mechanism
What does the system actually do?
Boundary
Where does the claim stop being true?
Test
What evidence would prove fit for this workflow?