Retrieval-augmented generation (RAG) is a method where an AI model first searches a defined set of documents, then writes its answer using what it found. In client campaigns, Alien Road uses RAG so that chatbots, content drafts and reports are based on the client’s own approved material rather than on the model’s general memory, which reduces invented facts and keeps claims consistent.
This article explains how RAG works, which campaign tasks it supports, and the process we follow from discovery to reporting.
How does RAG work?
A RAG system has three parts. A document store holds the source material, split into chunks. A retriever finds the chunks most relevant to a question, usually by comparing vector embeddings and often combined with keyword search. A language model then receives the question plus those chunks and writes an answer, ideally citing which chunks it used.
The model’s general knowledge still shapes the language, but the facts come from the retrieved text. If the retriever finds nothing relevant, a well-built system says so instead of guessing.
Why not just paste documents into a chat?
For a single question, pasting a document works. For a campaign with hundreds of product sheets, policy pages, past ads and research reports, it does not. Context windows have limits, costs grow with every token, and manual pasting means different team members get different answers.
RAG makes the source set fixed and shared. Everyone and every tool asks the same store, and when a document changes, you update it once.
Which campaign tasks use RAG?
- Website and WhatsApp chatbots that answer from product and policy documents
- Content drafts that must stay accurate to technical specifications
- Ad copy generation constrained by approved claims and compliance language
- Sales enablement assistants that find the right case study or answer for a rep
- Report commentary that explains performance using campaign notes and change logs
The common thread is that the output must match something the client has approved. RAG is less useful for open-ended creative work where you want the model to range freely.
How Alien Road applies retrieval-augmented generation
Discovery
A typical engagement starts with a source audit. We list the documents that hold the truth about the client’s products and claims: spec sheets, pricing rules, FAQs, legal disclaimers, brand guidelines. We also mark documents that are outdated or contradict each other, because RAG will faithfully repeat whatever is in the store, including old prices.
Setup
Documents are cleaned, split into chunks along natural sections such as headings, and tagged with metadata such as product, language, region and date. We index them in a vector database, commonly options such as pgvector, Pinecone or a managed search service, and pair it with keyword search so exact product codes are found reliably.
The prompt instructs the model to answer only from retrieved text, cite the source document, and state when information is missing. For multilingual clients, we index each language version rather than relying on translation at answer time.
Evaluation
Before launch, we build a test set of real questions with known correct answers, drawn from support tickets, sales emails and the client’s team. Each answer is checked for correctness, citation accuracy and refusal when it should refuse. We adjust chunk sizes, retrieval settings and prompts until results are stable.
Measurement and reporting
After launch, reporting covers answer accuracy on a reviewed sample, questions with no good retrieval (a sign that a document is missing), and the campaign metrics the system supports, such as leads from the chatbot or edit time on content drafts. The list of unanswered questions often becomes the next content plan.
What goes wrong with RAG?
- Stale documents: the store holds last year’s prices, so the bot quotes them
- Bad chunking: a table split across chunks loses its meaning
- Retrieval misses: the user’s wording differs from the document’s, and pure vector search does not find it
- Over-trust: teams assume RAG output is always correct and skip review
- Access control gaps: internal documents end up in a public bot’s store
Each of these is a process problem more than a model problem, which is why the audit and evaluation steps take more time than the technical setup.
Does RAG matter for SEO and AI search?
AI search products such as Google’s AI Overviews, Perplexity and ChatGPT search work on a similar principle: retrieve web pages, then generate an answer. Pages that are easy to retrieve and quote, with clear headings, direct answers and accurate facts, are more likely to be used. Building RAG systems for clients has shaped how we structure their public content, because we see which kinds of passages retrieve well.
RAG compared with fine-tuning
Fine-tuning changes a model’s weights by training it on examples. It is useful for teaching a style, a format or a narrow classification task. It is a poor way to teach facts that change, because every price update would need retraining, and the model may still blend new facts with old ones.
RAG keeps facts outside the model, in documents you can update in minutes. For campaign work, where products, offers and policies change monthly, RAG is almost always the better fit. The two can be combined: a model fine-tuned on brand tone, answering from a RAG store of current facts.
- Use RAG for facts that change: prices, stock, policies, specs
- Use fine-tuning for stable patterns: tone, output format, classification labels
- Use neither when a well-written prompt with a few examples already works
Who maintains the document store?
Every RAG system needs an owner on the client side who knows when a document changes. In a typical setup, product managers update specs, legal updates disclaimers, and marketing updates offers, each in their own folder. A weekly check flags documents older than a set age for review.
Where to start this week
Collect the ten documents your sales and support teams rely on most and check each for a date and owner. Remove or mark anything outdated. Then write twenty real customer questions with the correct answers next to them. Those two lists, clean sources and a test set, are the foundation of any RAG project, whatever tool you pick later.
Keep reading: AI Optimization · AI Optimization