What Is RAG (Retrieval-Augmented Generation)? A Guide

What RAG is #

RAG stands for Retrieval-Augmented Generation. It is the technique that lets a language model answer from your documents instead of from what it memorized during training. Before writing the answer, the system searches your own base (manuals, contracts, procedures, tickets, meeting notes) for the passages most relevant to the question and hands them to the model as context. The model writes from that evidence.

The idea solves the two problems that make a generic chatbot useless inside a company: it knows nothing about your operation, and when it does not know, it makes things up. With RAG the model has somewhere to look, and you can require it to cite where each statement came from.

The four stages of a RAG system #

1. Ingestion and chunking. Documents are converted to text and split into chunks of a few hundred words. Size matters: chunks that are too large dilute the search, chunks that are too small lose context.
2. Indexing. Each chunk is turned into a numeric vector (an embedding) that represents its meaning and is stored in a vector database. Many systems also keep a keyword index and combine both (hybrid search).
3. Retrieval. The user's question is turned into a vector, the closest chunks are retrieved and, often, a second model reorders them by relevance (re-ranking).
4. Generation. The selected chunks go into the prompt together with the question and the instructions, and the model writes the answer citing its sources.

RAG vs. fine-tuning #

They are complementary, but they solve different things. Fine-tuning retrains the model so it adopts a style, a format or a vocabulary; it is expensive, slow to update and does not ensure the model remembers a specific fact. RAG leaves the model untouched and gives it the data at query time: when a price or a procedure changes, updating the document is enough. For knowledge that changes, RAG wins. For behavior, fine-tuning.

The limits nobody mentions in the demo #

• Garbage in, garbage out. If the base holds two versions of a price list, the model will pick one with total confidence. RAG cannot tell current from outdated unless someone marks it.
• Chunks without context. A paragraph saying "the sensor runs on 24 V" may come from a different model than the one being asked about. Similarity search finds similar text, not correct text.
• Permissions. If the salesperson and the operator query the same base, the base has to know what each of them is allowed to see. Most demo RAG systems have no per-document access level.
• Questions that span sources. "Which motors on line 2 failed after we switched lubricant suppliers?" is not answered by one chunk; it is answered by walking relationships.

RAG with a knowledge graph #

The answer to that last limitation is to combine vector search with a knowledge graph: a structure where entities (machine, line, supplier, document, event) are connected by explicit relationships. It is usually called GraphRAG. Retrieval stops being "similar text" and becomes "the manual for the model installed on line 2, whose last work order was this one". On a plant floor, that is the difference between a plausible answer and a useful one.

RAG in an industrial company: what we learned #

At iGromi we built our own Second Brain so that people and agents could quote, support field work and answer customers with verified information. Three rules came out of that experience:

• Every piece of data carries four signals: who it belongs to, what access level it requires, what state it is in (unreviewed archive, draft, verified by a person, or in conflict with another source) and until when it is valid. An expired price is not delivered even if it exists.
• Empty is a valid answer. For prices, commercial terms and installation procedures, if the verified search returns nothing, the correct answer is "I don't have that information verified" plus a list of what is missing. It is never filled in by inference.
• No model reorganizes the memory on its own. AI can propose categories, relationships and summaries; a person approves them. That is what makes the base reliable for contracts and prices.

The same rules apply when RAG is connected to the plant: the agent that retrieves the installation procedure for a LoRa sensor gives nobody instructions without a source, and the one that proposes a work order does not execute it. We cover this in agentic AI in manufacturing and in industrial knowledge management.

How to start without ruining it #

1. Pick a small domain with real pain: prices and terms, or installation procedures.
2. Load only documents that have an owner and a date. Mark which version is current.
3. Define who can see what before connecting anyone.
4. Require citations in every answer and measure how many questions end in "I have no evidence". That number is your map of what still needs to be loaded.
5. Add the graph when questions start spanning sources, not before.

RAG FAQ #

What does RAG stand for?

RAG stands for Retrieval-Augmented Generation. Before answering, the language model searches a company's own document base for relevant passages and uses them as evidence to write the answer.

How is RAG different from fine-tuning?

Fine-tuning retrains the model on your data and blends it into its internal memory; it is expensive and hard to update. RAG leaves the model untouched and hands it the documents at query time, so updating the base is enough to change the answers.

Does RAG eliminate hallucinations?

It reduces them a lot, but it does not eliminate them. If retrieval returns a wrong or outdated passage, the model will use it with the same confidence. That is why the base has to be governed: verified sources, validity dates and permissions.

What is GraphRAG, or RAG with a knowledge graph?

It is RAG where retrieval does not just look for similar text but walks explicit relationships between entities: which machine belongs to which line, which manual matches which model. It answers questions that span several sources better.

Context: this article is part of the industrial artificial intelligence guide: what it is, where it applies and four cases in production in Chile.