Agentic RAG Node#
This is a RAG node where the agent explores documents by autonomously deciding the search strategy. During exploration, it selects folder and document scopes and combines multiple search methods to perform step-by-step exploration and reasoning.Unlike the standard Search (RAG) node, the Agentic RAG node allows the agent to determine the search strategy on its own in each round while exploring documents.
Instead of performing a single search with one query, it repeats exploration and evaluation over multiple rounds to collect the evidence needed for the answer.
It is suitable for cases where a single search is not enough to produce a sufficient answer, such as analytical, comparative, and synthesis-based questions that require cross-referencing multiple documents.
How It Works#
The Agentic RAG node internally operates as a multi-round exploration and reasoning pipeline.
In each round, the agent autonomously decides which search skills to use and what scope to explore.
The execution flow for each round is as follows:1.
Analyze the query and establish a search strategy
2.
Select the optimal search skill and perform exploration
3.
Extract and accumulate key facts from the retrieved chunks
4.
Determine whether the accumulated information is sufficient to answer the question
5.
If insufficient, adjust the strategy and proceed to the next round
Once sufficient evidence has been collected or a configured limit has been reached, the node synthesizes the collected results to generate the final answer.
Skills That Appear in the Reasoning Flow#
| Skill | What does it do? | When does it run? |
|---|
| 🧠 Previous Memory Retrieval | Loads previously accumulated search strategies and prior knowledge related to data from earlier conversations. | Runs once at the beginning of question processing. |
| 📁 File Search | Identifies documents and files related to the question and defines the search scope. | Runs when the scope of materials needs to be narrowed down. |
| 📄 Chunk Search | Searches within selected documents for passages that contain the answer to the question. | Runs when the details of a document need to be checked. |
| 🗄️ SQL Search | Queries the database directly to retrieve numerical data, aggregations, and tabular results. | Runs when statistical or quantitative information is required. |
| ✍️ Answer Generation | Synthesizes the collected materials to generate the final answer. | Runs last, after all search steps have been completed. |
Search Query#
This is the initial search query that the agent uses when starting document exploration.
It becomes the starting point of RAG exploration, and the first round of exploration is performed based on this query.
By default, {{query}} (user input) is set, but you can also define it directly using variables.
This is a required input field.
RAG Exploration Guide#
This is an optional input field that provides additional contextual information for the agent to reference during exploration.
If you provide guidance on the exploration strategy, domain knowledge, or the structure and scope of the learned documents, the agent can explore more accurately.
It works even if left empty, but providing it is recommended when the document volume is large or the domain is specialized.
Knowledge Scope to Search#
Specifies the scope of documents the agent will explore.
All documents learned by the agent: Explores all registered knowledge.
Documents learned in a specific folder: Explores only the documents within the selected folder.
Specifying a folder can reduce unnecessary document exploration and improve both accuracy and speed.
Maximum Exploration Rounds#
Sets the maximum number of times the agent can repeat document exploration and evaluation. (Range: 1–20, Default: 10)
In each round, the agent re-determines both the exploration target and the search method, and once the configured number of rounds is reached, it stops further exploration.
Increasing the number of rounds enables deeper exploration, but it also increases response time and token cost.
Maximum Allowed Exploration Time#
Sets the maximum time allowed for exploration. (Range: 0–1,200 seconds, Default: 180 seconds)
Even if all rounds are not completed within the allowed time, an answer is automatically generated based on the results collected up to that point.
In services with response time constraints, you can adjust this value to control response speed.
Number of History Turns#
Sets the number of previous conversation turns the agent will reference when generating a response. (Range: 0–5, Default: 5)
If you want to preserve context in multi-turn conversations, keep this value as is. If you want each question to be handled independently, set it to 0.
Exploration Termination Conditions#
Exploration in the Agentic RAG node ends when one of the following conditions is met.
| Termination Condition | Behavior |
|---|
| Sufficient evidence collected | The agent determines that the evidence is sufficient → Generates the final answer |
| Maximum rounds reached | The configured number of rounds is reached → Generates an answer based on the results collected so far |
| Maximum time exceeded | The configured allowed time is exceeded → Generates an answer based on the results collected so far |
You can check which condition caused termination in the Reasoning Process tab of the execution log.
Viewing the Execution Log#
In the execution log of the Agentic RAG node, you can review the agent’s reasoning process in detail.
Reasoning Process tab: Displays, in chronological order, the search skills used in each round, the number of retrieved chunks, the number of extracted facts, token usage, and confidence scores.
Detailed Log tab: Lets you review the final answer generation result and the full execution data in JSON format.
Example Use Cases#
Answering questions that require comprehensive analysis by cross-referencing multiple documents
Searching large volumes of documents to precisely find specific information
Handling complex context-based queries that are difficult to resolve through simple keyword matching
Building comprehensive answers by combining internal documents with web search
If the Agentic RAG node is executed in parallel within a workflow, token costs can increase significantly.
If a parallel configuration is necessary, we recommend setting the exploration rounds and allowed time conservatively.
