Retrieval-Augmented Generation (RAG)
Definition
RAG connects a language model with an external, searchable knowledge base so that answers are based on retrieved information and not only on the model parameters.
RAG models combine a pretrained language model (parametric memory) with an external knowledge base (non-parametric memory) accessed by a neural retriever. In the original paper by Lewis et al., this knowledge base is a dense vector index of Wikipedia. For retrieval, the Azure AI Search documentation mentions vector, keyword, and hybrid search.
Simply explained
Like looking up before answering: You first search for relevant text passages and then formulate the answer from the found information.
Why it matters in practice
Enterprises can link language model answers to their own documents. For this, content must be prepared and access rights implemented so that only authorized information is retrieved.
Example from business
Possible example: Service technicians of a mechanical engineering company search through a large collection of PDF maintenance manuals via chat during customer conversations.
Limits and pitfalls
- Unreadable scans, unsuitable text sections, and outdated search indexes degrade retrieval quality. Missing access controls can expose confidential content; manipulated documents may contain hidden instructions.
Origin & history
Lewis and other authors submitted the work on RAG for knowledge-intensive language tasks as a preprint on May 22, 2020; it was accepted at NeurIPS 2020.
Use in mid-sized companies
- Possible applications: internal knowledge management, assisting ticket responses in support, and support in contract review with professional result verification.
Frequently asked questions
Does RAG automatically change model weights?
No. During retrieval it uses search results; additional fine-tuning is a separate step.
Does RAG help with outdated model knowledge?
An updated knowledge base can provide more recent information.
Do the source passages for an answer remain the same?
Depending on the variant, the same passages are used for the entire answer or different passages are used per token.
Sources
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks · Lewis et al. (2021): Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- Retrieval-augmented generation (RAG) in Azure AI Search · Microsoft (2026): Retrieval-augmented generation (RAG) in Azure AI Search
- Hybrid search using vectors and full-text search in Azure AI Search · Microsoft (2026): Hybrid search using vectors and full-text search in Azure AI Search
- Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection · Greshake et al. (2023): Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
Sources checked on 14.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on:
On this page
Next step
Use RAG in your own company?
Together, we look at where your knowledge is stored today and how an assistant can use it to produce reliable, traceable answers.
- Assess which documents and data sources are suitable
- Clear information on privacy and access rights
- No obligation and no sales pressure