Retrieval-augmented generation (RAG)

The AI first looks in an approved library, then uses what it finds to help answer. This makes answers easier to update and support with sources.

Where it creates value

Best suited to work like this.

Policy and procedure assistants, scientific knowledge support, document question answering, and source-grounded drafting.

What it does

The intelligent task.

Ground generation in selected documents, retrieve current knowledge, attach citations, and limit answers to an approved corpus.

Typical value

Makes generative answers more relevant, current, and traceable by grounding them in governed organizational knowledge.

Where it performs well: Knowledge can be updated without retraining the model and answers can point to retrievable sources.

Value maturity · editorial assessment Scaling

Retrieval-augmented generation is a common production pattern for grounding model responses in governed knowledge, with scaling constrained mainly by source quality, evaluation, and access controls.

Emerging
Demonstrated
Scaling
Established

Back

TEC-0004 · How it works in practice

Retrieval-augmented generation (RAG)

Read the card type together with the use case. A component can be central in one solution and enabling in another.

01 · Inputs

User question, indexed documents, metadata, embeddings or search terms, retrieval rules, and generation instructions.

02 · Intelligent task

Ground generation in selected documents, retrieve current knowledge, attach citations, and limit answers to an approved corpus.

03 · Outputs

A generated answer or draft with retrieved passages, references, confidence cues, or abstention.

Common application patterns

Policy and procedure assistants
scientific knowledge support
document question answering
and source-grounded drafting

Enablers

What must be in place.

Curated source content, indexing and retrieval, permissions, chunking and metadata design, model integration, evaluation, and citation handling.

Where it struggles

Limitations and drawbacks.

Bad retrieval, poor document segmentation, outdated sources, or irrelevant context can still produce incorrect or unsupported answers.

Human role

People remain accountable.

Knowledge owners maintain the source collection; users verify cited evidence and do not treat retrieval as proof that an answer is correct.

Connected use cases

Examples from the current use-case library.

Reduce document workloadAutomate regulatory document preparationRegulatory · Pilot Find safety signals earlierDetect adverse event signals in literatureSafety · Pilot

Best suited to work like this.

The intelligent task.

What must be in place.

Limitations and drawbacks.

People remain accountable.

The card is the simple front door. The website can hold the full system behind it.