Vector database
Search the workspace's collections and return the most relevant passages.
The Vector database runs a semantic search over the workspace's Collections and returns the passages closest to the query.
It is the node that enables RAG inside a Workflow, the equivalent of attaching Collections to an Agent.
Configure
The configuration has three parts, and each answers a different question:
- Collections: where to search. One or several, out of the workspace's.
- Search text: what to search for. Usually the question coming from the Input, but the field accepts a composed query, such as
{{ entrada.pergunta }} in the context of {{ entrada.categoria }}. - Number of chunks: how much to bring back, through a slider.
How many chunks to bring back
| Range | When to use it |
|---|---|
| 3 to 5 | Objective questions. Balances context and cost. |
| 5 to 10 | Broad summaries and research. More context, more tokens in the next LLM node. |
Raising the number doesn't always help: barely relevant passages dilute the context and increase the chance that the model leans on the wrong material.
This node requires at least one configured Collection. An empty list doesn't mean "search everything". It means the node has nowhere to look, and it returns nothing.
What it returns
The output is chunks: a list with the texts found, the similarity score and the source metadata.
It doesn't write an answer. The LLM node that follows is what turns the passages into text, receiving chunks in its prompt.
When it returns nothing
Three causes, in order of frequency: the Collection's documents didn't reach the Active state, the configured Collection is empty, or the search text has nothing to do with the indexed content.
The run log shows what the node returned, which tells the third case apart from the first two.