Retrieval-augmented generation (RAG) workflows save time by letting a language model answer from your own documents instead of from memory. A RAG system searches a knowledge base for relevant passages, passes them to the model with the question, and asks for an answer grounded in those passages, ideally with citations. For marketing teams, it turns brand guidelines, product specs, past campaigns and research into something people can query in seconds.
What problem does RAG solve?
Language models know a lot about the world and nothing about your company’s latest price list. Fine-tuning a model on internal documents is slow and does not handle frequent updates well. RAG keeps knowledge outside the model, in a searchable store, so updating a document updates the answers.
It also reduces made-up answers. When the model is told to answer only from retrieved passages and to say when nothing relevant was found, it invents less. Not never, which is why evaluation matters.
What are the parts of a RAG workflow?
- Document collection: the files and pages the system should know.
- Preparation: cleaning, converting to text, removing duplicates and outdated versions.
- Chunking: splitting documents into passages small enough to retrieve precisely.
- Embedding: turning each passage into a vector with an embedding model.
- Storage: a vector database or search index, such as pgvector, Pinecone, Weaviate, Elasticsearch or Azure AI Search.
- Retrieval: finding the passages most relevant to a question, often mixing keyword and vector search.
- Generation: the model writes the answer from those passages, with references.
Managed services bundle many of these steps. OpenAI’s file search in the Assistants and Responses APIs, Google’s Vertex AI Search, Amazon Bedrock Knowledge Bases and Azure AI Search with Azure OpenAI all handle chunking and retrieval for you. For a first project, a managed option is usually faster.
A step-by-step workflow for a marketing knowledge base
Step 1: Choose a narrow scope
Start with one question type, such as “what can we say about product X in ads?”. Collect only the documents that answer it: product specs, approved claims, brand guidelines and legal notes. A small, clean collection beats a huge messy one.
Step 2: Clean the documents
Remove old versions, or at least label them with dates. Convert PDFs to text and check tables came through correctly. Scanned documents need OCR. This step is dull and decides most of the quality.
Step 3: Chunk by meaning
Split by headings or sections rather than by a fixed number of characters where possible. Keep each chunk focused on one idea, and attach metadata: document title, section, date, product and language. Metadata lets you filter retrieval, for example to only current documents.
Step 4: Retrieve with hybrid search
Pure vector search misses exact terms like product codes. Pure keyword search misses paraphrases. Most production systems combine both and then re-rank the top results with a re-ranking model. Retrieve more than you need, then keep the best few.
Step 5: Prompt for grounded answers
Give the model clear rules: answer only from the provided passages, cite the passage title for each claim, say “not found in the documents” when the answer is missing, and never guess prices or dates. Put the rules in the system prompt so users cannot remove them by accident.
Step 6: Show sources to the user
Display the cited passages next to the answer with links to the original file. People trust the answer more and catch errors faster when the source is one click away.
How do you know the RAG system works?
Build a test set of 50 to 100 real questions with known correct answers and the document each answer comes from. Measure two things separately. Retrieval quality: did the right passage appear in the top results? Answer quality: was the final answer correct and supported by the passage?
Splitting the two shows where to fix problems. If retrieval fails, adjust chunking, metadata or search settings. If retrieval works but answers are wrong, adjust the prompt or the model. Tools such as Ragas and similar open-source evaluators can automate parts of this, but have people check a sample.
How Alien Road applies RAG in client work
In a typical engagement our AI optimization team starts by interviewing the people who answer internal questions most often, for example the brand manager or product marketer, and collecting the questions they get each week. Those questions become the test set.
We then gather and clean the source documents, choose a managed retrieval service or a custom pipeline depending on data rules, and build a simple interface inside a tool the client already uses, such as Teams, Slack or the intranet. Answers always show sources.
Measurement covers retrieval hit rate and answer accuracy on the test set, weekly usage, and questions that returned “not found”, which point to missing documents. Monthly reports list document updates made and the change in scores.
Which mistakes slow RAG projects down?
Most delays come from the same few habits. Teams load every document they own on day one and then cannot tell why answers are poor. They skip the test set and argue about impressions. They pick a vector database before knowing how many documents they have, when a managed service would do. And they forget permissions: if a sales deck is restricted in SharePoint, the RAG system should respect that restriction, or confidential pricing ends up in an answer for the wrong person.
- Start with one collection, not all of them.
- Write the test set before choosing tools.
- Carry document permissions into retrieval filters.
- Log every question and answer for review.
How do you keep it current?
RAG systems decay when documents change and nobody re-indexes them. Set up automatic syncing from the source folder, such as SharePoint or Google Drive, and give each collection an owner who reviews it quarterly. Remove expired campaigns and old price sheets rather than leaving them for the retriever to find.
Where to start this week
List the ten questions your team answers repeatedly from documents. Collect the files that contain the answers into one folder, remove outdated versions, and try a managed file search option with those ten questions. If at least seven answers come back correct with the right source, you have a scope worth building on.
Keep reading: AI Optimization · AI Optimization