Here are the nine rag patterns you need

to know if you're serious about building

AI systems. Let's start with our

deterministic designs. Number one is

naive rag. This really is the baseline

of rag architectures. And while it can

work for simple FAQs style scenarios,

you'll usually need more than this when

you're in a production setting. Number

two is query transformation with rag

fusion. Here we are decomposing and then

expanding on a user's query to fetch a

wide range of chunks. Number three is

iterative retrieval. Unlike the previous

design, here we have multiple passes for

complex questions. Next up is adaptive

retrieval. Here we can decide whether we

need to retrieve at all. And if we do,

is it simple or is a multi-stage

retrieval required? And now onto our

agentic designs, which brings us to

number five, agentic rag. Here we are

providing the LLM with decisionmaking

powers. It can decide what to retrieve

and how many times to carry out

retrieval before answering the question.

And a variant of this is hybrid rag

where the agent can pull from knowledge

graphs, SQL databases as well as vector

stores. And then we get to the realm of

multi- aents where we have a main agent

and sub agents. And here we can

distribute the cognitive load as well as

protect the context window of the main

agent so we can answer the user's

question. Instead of having AI agents

sitting under a supervisor agent, you

can have sequential agents. This allows

you to have multiple specialist AI

agents in sequence. And finally, we get

to agentic routing. Here, we can

intelligently route the query to the

best agent for the job. There's no

one-sizefits-all with Drag, so you'll

need to choose the best pattern for your

use