ChromaDB
On this page
Purpose
ChromaDB stores embedded Data and searches it by semantic similarity. It can return matches to a flow or expose a search tool to Agent Core.
When to use it
- Build retrieval-augmented responses from prepared Data.
- Search a document collection by meaning.
- Give an agent access to a bounded collection of indexed content.
Required setup
Connect an Embedding Model to Embedding Model. Connect Data or DataFrame values to Ingest Data only when the node should add records.
How to use it
Tool Mode
- Index the collection with a Normal mode ChromaDB node.
- Add a second ChromaDB node with the same collection and embedding settings.
- Enable Tool Mode and connect Toolset to Agent Core.
- Test a specific retrieval question before publishing.
Normal mode
- Prepare source content as Data and connect it to Ingest Data.
- Connect an Embedding Model.
- Use a separate query node with the same collection and embedding settings.
- Enter Search Query and use Search Results or the DataFrame output.
Configuration
| Setting | What it controls | Recommended starting point |
|---|---|---|
| Collection Name | Base name for the isolated collection. | Use a short descriptive name. The default is arkus. |
| Ingest Data | Data or DataFrame records to add to the collection. | Connect it only on the node that should add records. |
| Embedding Model | The connected embedding model used to embed and search records. | Connect the same Embedding Model used to index the collection. |
| Search Query | Text used for semantic search. In Tool Mode, this is the agent argument. | Start with a specific question. |
| Cache Vector Store | Reuses the vector store during one component execution. | Leave on. It does not persist across runs. |
| Allow Duplicates | Checks an existing-record window before ingestion. | Leave off, but do not rely on it to remove duplicates within one incoming batch. |
| Search Type | Selects Similarity or MMR retrieval. | Start with Similarity. |
| Number of Results | Limits returned matches. | Start with 10. |
| Limit | Limits records checked for duplicate comparison. | Leave at 0 unless a bounded comparison is needed. |
Expected result
Search Results returns a list of Data records. The DataFrame output wraps the same matches. Results do not include similarity scores, and complex metadata can be removed before ingestion.
Reference details
Arkus provides the store itself, so there is no server, port, or path for you to set. Use separate ingestion and query nodes when the flow needs predictable retrieval behavior.
Publishing copies the collection for that version, and the published flow only reads that copy. It can copy content that Ingest Data draws from Knowledge Base - Files, Structured File, or Web Fetch, including through Split Text, because Arkus can gather that content at publish time. Publishing stops when Ingest Data comes from anywhere else, because that content would have to be written while the published flow runs. Ingesting again in Builder afterwards does not change what the published flow reads.
If the embedding connection is missing, reconnect Embedding Model and rebuild. Save the flow before running the node: each flow keeps its own separate collection.
Related components
- Embedding Model supplies the required embeddings client.
- Agent Core can use ChromaDB in Tool Mode.
- Data Operations can shape Data before ingestion.