What Is RAG (Retrieval-Augmented Generation)?
The Problem
Section titled “The Problem”An LLM only knows its training data, not a customer’s internal company documents. Without a connection to current or company-specific sources, hallucinations or outdated answers result.
The RAG Solution
Section titled “The RAG Solution”RAG combines retrieval (search) with generation (answer creation) in three steps:
- Retrieval: embeddings are used to find semantically relevant passages from a vector database.
- Reranking (optional): a reranker sorts the hits by true relevance.
- LLM request with context: the question plus the retrieved context is sent to the LLM, which generates the answer from it.
Benefits
Section titled “Benefits”- Current knowledge without retraining
- Traceable sources
- Cheaper than fine-tuning
noris Models for RAG
Section titled “noris Models for RAG”| RAG Component | noris Model | Task |
|---|---|---|
| Embedding | Harrier OSS v1 0.6B | Text → vectors for semantic search |
| Reranker | BGE Reranker v2 M3, Jina Reranker v2 | Sorts results by relevance |
| LLM (Generation) | Gemma 4, GPT-OSS, GLM 5.2, Qwen 3.6 | Generates the answer from question + context |
A practical step-by-step guide is available under RAG Pipeline.