What's a RAG AI in ARPIA?

A RAG AI answers questions using content retrieved from your organization's documents and data. Instead of relying only on what a language model learned in training, it looks up relevant information at query time and builds its answer from that.

Use a RAG AI when answers need to be grounded in documents and knowledge sources. If people mainly need to query structured data, use an AI Agent. If you need to run a repeatable process, use an AI Worker.

What is RAG?

Retrieval-Augmented Generation (RAG) adds a retrieval step before the model generates a response.

A language model on its own:

Only knows what it learned in training, which goes out of date
Has no knowledge of your organization's policies, procedures, or data
Tends to give general answers

RAG closes that gap by retrieving relevant content first and using it as the basis for the answer.

What it does

Takes a question in natural language
Finds relevant content in sources the user is authorized to access
Generates an answer based on that content
Returns a summary, comparison, key findings, or a visualization (when configured)

How it works

Ask. The user asks a question in natural language.
Retrieve. The system finds relevant content in authorized sources.
Generate. The language model writes a response using the retrieved content as context.
Respond. The answer is returned as text, structured data, or a visualization.

Data sources

A RAG AI retrieves from authorized ARPIA components, including:

Data Models (Nodes)
Structured datasets in the platform
Embedded documents
Connected knowledge repositories

Example uses

Use case	Example
Knowledge search	Find the current policy, procedure, or operating guideline
Compliance support	Retrieve the documentation that supports a control or requirement
Data summaries	Summarize trends from structured data alongside related documents
Decision support	Feed retrieved context into a larger workflow

Access and governance

Role-based access. Retrieval only covers content the user's Security Profile allows.
Read-only by default. Write access requires explicit configuration.
Scoped by workflow. Retrieval follows the permissions and settings in the configured workflow.

For governance that applies to all AI implementations (human review, logging, change management), see AI-Driven Implementations.

Limits

Grounding reduces errors but doesn't eliminate them. The model can still misread or misstate retrieved content.
Answers are only as current as the sources. Outdated or missing documents produce outdated or incomplete answers.
Retrieval can miss relevant content. It depends on how sources are prepared and how the question is phrased.
[VERIFY: do responses show which sources were used?] If so, say so here and tell readers to check them for high-stakes decisions.

What's a Rag AI in ARPIA?