# Agentic mode: coreAI builds its own context when the answer is missing

When coreAI lacks coverage in the context it has already been given, agentic mode lets the assistant build the context itself – instead of guessing or giving up. It picks its own tools along the way: semantic search, keyword search and structured filtering against your own knowledge base – plus external MCP APIs if you have connected the assistant to relevant sources. You enable the feature per assistant, and it changes how it thinks, not just what it knows.

![Sketch of how coreAI selects tools in agentic mode](/articles/agentic-mode.webp)

## What is agentic mode?

Traditional RAG (Retrieval-Augmented Generation) loads context up front: the system runs one search, pushes the results into the prompt, and lets the language model answer from what it was given. Agentic mode inverts this. Instead of pre-loading context, knowledge base search is exposed as a tool the language model can call itself – as many times as it needs to cover the question.

That means the assistant makes its own decisions along the way: which searches to run, how to phrase them, and when it has enough to answer. On multi-part questions, that is the difference between half an answer and a whole one.

## How the assistant reasons in several steps

After each search coreAI runs a coverage check: it identifies the distinct sub-questions and named entities in what the user asked, and verifies that each one has direct support in the sources that came back. If something is missing, it searches again with a broader or sharper query before it even begins to answer.

![Sketch of how coreAI runs a coverage check and searches again before answering](/articles/agentic-mode-reasoning.webp)

- on substantive questions the assistant is required to search rather than guess
- it can run several searches within a single answer to catch every part of a question
- if it still finds no support, it says the answer isn't in the knowledge base – instead of inventing one
- all hits from the searches are accumulated, so the answer can cite exactly which sources it rests on

Example: "What is your return policy, and how long does a refund take?" is really two questions. Agentic mode searches first for the return policy, then for the refund timeline, and answers both. A single pre-load would easily miss one half.

## The tools the assistant can use

Agentic mode gives the language model a small, precise set of tools – and it picks the right one itself:

- semantic search for natural-language questions, broad topics and recommendations
- keyword search for looking up exact terms: person names, roles, document titles or product numbers
- structured filtering for product assistants, for example on price, product number or whether an item is on sale
- external tools via MCP (Model Context Protocol) – for live data or actions that sit outside the knowledge base

When an assistant has both its own knowledge base and connected MCP servers, it can combine internal and external tools in the same answer: pull facts from your own sources and check a live status from another system in one and the same turn.

## When agentic mode makes the biggest difference

Agentic mode pays off most when questions are complex, when the answer is spread across several sources, or when the user expects the assistant to fetch something fresh from another system. You get more complete answers, verifiable source citations, and an assistant that would rather say so than hallucinate when the information isn't there.