In words
What it is, why it matters, and what it is like.
Why am I learning this?
This unlocks the ability to build retrieval-augmented generation (RAG) systems that can answer questions no single search can answer. You'll be able to handle questions that require multiple pieces of evidence in sequence, recover from a bad first guess, and keep costs under control. This is the difference between a chatbot that gives up after one search and an AI assistant that keeps digging until it finds the answer. It leads directly to studying AI Agents, Multi-Document Retrieval, and Hallucination Mitigation.
The idea, in plain terms
Imagine you're a detective. A witness gives you a vague clue: 'The suspect was seen near a restaurant that's famous for its biryani.' You start your search with that clue. You find a list of biryani restaurants, but you need more: which ones are near the crime scene? You check the map, narrow it down. Now you have a shortlist, but you need the suspect's name. You interview staff at each restaurant. 'Did you see anyone suspicious?' One says, 'Yes, a man in a red jacket.' Now you search for 'man in red jacket' in the records. You find a match. This is not a single search; it's a chain of searches, each step guided by the result of the last. Agentic retrieval is exactly this: instead of one query-and-done, the model decides what to search for, looks at the results, decides if it has enough to answer, and if not, searches again with a refined query. It's a loop of 'search, evaluate, refine' until it's confident.
An analogy
Think of a chef preparing a dish from a complex recipe. The recipe says 'add spices to taste.' The chef doesn't just add a fixed amount; they add a little, taste it, evaluate if it's too bland, too salty, or just right. If it's not right, they adjust the spice, taste again, and repeat until it meets their standard. Agentic retrieval is the same. The 'query' is the initial spice. The 'taste' is the model evaluating whether the retrieved documents contain the answer. The 'adjustment' is the next refined query. This loop is natural for a chef, but for a computer program, it's a novel idea because traditional programming is a straight line: input -> process -> output. No loop back. The analogy breaks down when you consider that the chef is a single, integrated intelligence; in AI, the 'chef' is a language model that can both generate queries and evaluate results, but it has a limited 'taste'—it can only see a certain number of documents at once, and it might be wrong about what it's tasting. So unlike a chef, the model needs a strict limit on how many times it can taste, or it will keep cooking forever, burning the budget.
Definition
Agentic retrieval is a retrieval-augmented generation strategy where the language model iteratively generates search queries, evaluates the retrieved results against its need for information, and decides to either generate the final answer or perform another search, repeating until it satisfies a stopping condition.
Where this sits
This concept builds directly on Retrieval-Augmented Generation (RAG). In a standard RAG pipeline, you have a single query: you embed it, search your vector database, get a few chunks, and generate. Agentic retrieval is a _loop_ around that pipeline. It also connects to your notes on Retrieval Quality and Context Precision and Recall because the model needs a way to evaluate if the retrieved chunks are relevant (precision) and if it has all the needed chunks (recall). It relates to Multi-Document Retrieval because in each iteration, the model might assemble evidence from multiple documents, and it relates to Chunking because if your chunks are too small or too large, the model will struggle to find the answer and will need more iterations to piece things together.
