Chroma

Open-source AI-native vector database for agent memory. Manage collections and records, and run nearest-neighbor similarity search against Chroma Cloud or a self-hosted server for RAG retrieval.

12 actions available

New source text is ready to enter the knowledge base

Agent reads the text and its embedding from an upstream step

Agent writes with Upsert Records into the target collection

Agent confirms the write with Count Records

Agent runs Query Records to confirm the content is retrievable

Agent reports the collection size change to the owning team

Any Delete Collection call waits for an owner to approve

What This Integration Enables

The Chroma connector exposes collection management and the full record lifecycle. Agents create, list, get, delete, and count collections, then add, upsert, update, query, get, count, and delete records. Add Records, Upsert Records, and Update Records take positionally aligned arrays of IDs, embeddings, documents, and metadata. Query Records runs nearest-neighbor search with metadata filtering (where) and document-content filtering (where_document), and controls which fields come back. Record operations accept either a collection name or its UUID, so an agent does not have to resolve the ID first. The same actions work against Chroma Cloud and a self-hosted instance. Agents usually embed source text with a provider such as [OpenAI](/integrations/openai-ai) or [Cohere](/integrations/cohere) before an Add Records call, then log retrieved matches into a system like [Notion](/integrations/notion).

Without FlowRunner

No place to search Reference text sits in files no agent can query by meaning

Store falls behind Embeddings go stale as the source content moves on

Coarse retrieval Search returns matches with no way to scope by metadata

With FlowRunner

Semantic recall Query Records returns the closest chunks for an LLM prompt

Store stays current Upsert Records keeps embeddings in step with source content

Scoped retrieval Metadata and document filters narrow results before ranking

Use Case Scenarios

RAG Knowledge Base

A team wants an assistant grounded in its documentation. The agent embeds each chunk with an external model and stores it with Add Records, carrying the source path and section as metadata. On a user question the agent embeds the query, calls Query Records with a where filter that scopes to the relevant product area, and hands the top chunks to an LLM. The metadata filter keeps retrieval tight even as the collection grows.

Keeping The Store In Sync

Source content changes weekly. When a document is updated, the agent re-embeds the affected chunks and calls Upsert Records, matched by ID, so existing entries are replaced in place rather than duplicated. Count Records confirms the collection size after the write. The store stays a faithful mirror of the source without a rebuild.

Retiring A Collection

A collection is no longer needed. The agent reads its record count with Count Records and its recent query activity, assembles a removal summary, and routes it to the data owner. Only after approval does it call Delete Collection, which is permanent.

Human-in-Loop Highlight

Querying a collection is safe. Deleting one removes every record in it for good. [Human-in-the-loop](/concepts/human-in-the-loop/) is an execution pattern where AI agents pause autonomously, assemble the relevant context and the decision choices available, route to a human via their preferred channel, and resume the moment the human responds. The Chroma connector places that pause on Delete Collection. When a flow reaches that step, the agent first pulls the record count with Count Records and the last query activity, then asks the owner through their channel: "The legacy-handbook collection holds 6,200 records and was last queried three weeks ago. Deleting it is permanent. Delete it?" The agent handles ingestion, upserts, and search on its own. A person owns the decision to erase a collection the team may still rely on.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

12 actions

Collections

Create Collection Create a new collection in the configured tenant and database, optionally with metadata and index configuration. Returns the collection UUID needed by record operations.
List Collections List the collections in the configured tenant and database with pagination.
Get Collection Retrieve a single collection by name or UUID, returning its id, name, metadata, and configuration.
Delete Collection Permanently delete a collection and all records stored in it, identified by name. This cannot be undone.
Count Collections Return the total number of collections in the configured tenant and database.

Records

Add Records Add records to a collection: record IDs plus embeddings, documents, and metadata, aligned positionally by index.
Upsert Records Insert new records or update existing ones matched by ID. Use it to keep an embedding store in sync as source content changes.
Update Records Update existing records matched by ID, changing only the fields provided. IDs that do not exist are ignored.
Query Records Run a nearest-neighbor similarity search using one or more query embeddings, with optional metadata and document filtering. The core retrieval operation for RAG.
Get Records Retrieve records by explicit IDs and/or metadata and document filters, without similarity scoring.
Count Records Return the total number of records stored in a collection.
Delete Records Delete records from a collection by explicit IDs and/or metadata and document filters.

Start building with Chroma

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us