Tags let one knowledge base serve several brands, products or regions at once. You attach tags to the content you upload, then filter retrieval to one combination of them in a flow, so the AI agent answers from the right slice of the knowledge base instead of all of it. This is an alternative to splitting the same content across separate knowledge bases.
For creating a knowledge base and adding data sources to it, see Creating a Knowledge Base, Adding Documents and Adding a Website.
📝 Note: Tags are defined per knowledge base, and nothing else works until at least one exists. A source row offers no tags to pick, and a flow node has nothing to filter on, until the knowledge base defines them.
Defining tags
A tag belongs to one knowledge base and has a name and a description. The Tags section appears in two places, and they are the same section:
- As the fifth step of the create knowledge base wizard.
- As a section in the settings of a knowledge base that already exists.
So you do not have to decide your tags up front. Add them at any point, and tag sources as you upload them.
1. Name: What the tag is called. This is the name you will select when tagging a source, and the name an FAQ spreadsheet column heading has to match exactly. It starts with a letter or an underscore and can then contain letters, digits and underscores only, so brand and product_line are accepted and product line is not. Up to 64 characters.
2. Description: What the tag is for. It is there for whoever maintains the knowledge base after you, and does not affect retrieval.
The examples below come from a knowledge base holding support content for two appliance brands, Vantera and Solmar. It defines two tags:
brand Which appliance brand the content belongs to
product Which appliance the content is about
Tagging a website or a document
On the Add Data Source screen, each website row and each document row carries its own Add tag control.
1. Add tag: Opens the list of tags the knowledge base defines. A tag you have already used on this row is not offered again, so each tag appears at most once per row. The control is disabled once the row uses every tag the knowledge base defines.
2. Tag name: The tag you picked, for example brand.
3. Value: What you type for this row, for example vantera. The value is free text and belongs to this row alone, so the next document can carry brand: solmar under the same tag.
Tag each row before you save the source. A manual tagged brand: vantera and product: refrigerator is then reachable by a flow filtering on either or both of those.
Tagging FAQ rows
FAQ Data works differently from websites and documents, and this is the part most likely to catch you out: the FAQ Data screen has no tag picker. The tags come from the spreadsheet itself.
The file's columns are read like this:
| Column | What it holds |
|---|---|
| First column | Question. |
| Second column | Answer. |
| Every further column | One tag. The heading must match the name of a tag defined in this knowledge base, and each cell holds that row's value for that tag. |
An empty cell simply means that row carries no value for that tag, so a file can mix rows that are brand-specific with rows that apply to every brand.
Using the two tags above, a file for this knowledge base looks like this:
Question | Answer | brand | product
How do I reset the refrigerator? | Hold the Mode button for ... | vantera | refrigerator
How long is the warranty? | Two years from purchase ... | vantera |
💡 Tip: Download the template the FAQ Data screen offers. It is built from this knowledge base's own tags, so the headings come out spelled the way the tags are spelled, which is what the match depends on.
Checking what a source carries
The Data Sources list shows the tags on everything you have uploaded, and it shows the two kinds of source differently.
1. A document or website row: Lists each tag as name and value together, because the whole source carries one value per tag.
2. An FAQ row: Lists the tag names only. The values come from the spreadsheet and differ from row to row, so there is no single value to show for the file.
Filtering retrieval in a flow
The filter lives on the flow, not on the content. Both the Knowledge Retrieval node and the RAG Agent node carry a Tags section in the flow designer.
1. Tag: One of the tags defined on the knowledge base the node queries.
2. Value: What that tag has to equal for content to be returned.
Add one row per tag you want to filter on. Retrieval returns only content matching all the tags you give, so brand: vantera plus product: refrigerator returns the content carrying both.
A value does not have to be fixed. It can also be taken from a conversation variable, which is how a single flow serves both brands: identify the caller's brand earlier in the conversation, write it to a variable, and filter on that variable here.
Make sure the value the variable holds matches the value on the content, including its spelling. For conversation variables, see Variable Management.
📝 Note: The tag filter is available on the Knowledge Retrieval and RAG Agent flow nodes. An AI agent's Knowledge Retrieval tool does not have one in this release, so retrieval that has to be narrowed by tag belongs in a flow.
Editing and deleting a tag
A tag that is in use cannot be edited or deleted. Remove it from the data sources and the flow nodes using it first, then change or delete the tag itself.
What an unfiltered retrieval does
A knowledge base holding two brands does not notice that it holds two brands. Asked a question both brands answer differently, retrieval returns whatever matches best and the AI agent answers from it, by name and with confidence. It does not hedge, and it does not ask which brand the customer owns.
In our own testing, a question about resetting a refrigerator came back with the Vantera procedure, named as such, to a customer who might have owned the Solmar. Nothing on the screen suggested the answer was one of two.
That is what the filter prevents. If a knowledge base carries content for more than one brand, product or region, tag it and filter on it.