<a id="cc-chromadb-sink"></a>

# ChromaDB Sink Connector for Confluent Cloud

The fully managed ChromaDB Sink connector for Confluent Cloud streams data from Apache Kafka®
topics into ChromaDB collections, with optional built-in embedding generation
for retrieval-augmented generation (RAG), semantic search, and recommendation
use cases.

[ChromaDB](https://www.trychroma.com/) is an open-source vector database for
artificial intelligence (AI) and machine learning (ML) workloads. The
connector writes records to one or more ChromaDB collections and, when
needed, generates the vector embeddings for you by calling an
OpenAI-compatible embedding API.

## Features

The ChromaDB Sink connector provides the following features:

* **Embedding generation**: Optionally generates vector embeddings from
  document text by calling an OpenAI-compatible embedding API (for example,
  OpenAI, Azure OpenAI, or another compatible provider). When disabled,
  embeddings must already be present in each Kafka record.
* **Pre-computed embeddings**: Supports records that already contain a
  computed embedding, so you can skip the embedding-generation step entirely.
* **Collection auto-creation**: Automatically creates a ChromaDB collection if
  it does not already exist.
* **Input formats**: Supports Struct, Map, and JSON String record
  formats.
* **Multi-tenancy**: Supports configurable tenant, database, and collection
  settings for both Chroma Cloud and self-hosted ChromaDB deployments.
* **Token-based authentication**: Authenticates to ChromaDB using an API
  token.
* **Multi-collection support**: A single connector can write to up to 15
  ChromaDB collections, mapping each Kafka topic to a different collection.

For more information and examples to use with the Confluent Cloud API for Connect,
see the [Confluent Cloud API for Connect Usage Examples](connect-api-section.md#ccloud-connect-api) section.

## Limitations

* For connector limitations, see [ChromaDB Sink Connector](limits.md#cc-chromadb-sink-limits) limitations.
* If you plan to use one or more Single Message Transforms (SMTs), see
  [SMT Limitations](single-message-transforms.md#cc-single-message-transforms-limitations).

## Quick Start

Select the connector and configure it to stream Kafka events to a ChromaDB collection.

<a id="cc-chromadb-sink-prereqs"></a>

### Prerequisites

- Authorized access to a
  [Confluent Cloud](https://www.confluent.io/confluent-cloud/) cluster
  on Amazon Web Services (AWS), Microsoft Azure (Azure), or Google Cloud.
- The Confluent CLI installed and configured for the cluster. For more
  information, see
  [Install the Confluent CLI](https://docs.confluent.io/confluent-cli/current/install.html).
- [Schema Registry](../get-started/schema-registry.md#cloud-sr-config) must be enabled to use a
  Schema Registry-based format (for example, Avro, JSON_SR (JSON Schema), or
  Protobuf).
- A Chroma target: either a Chroma Cloud account or a self-hosted ChromaDB
  endpoint reachable from Confluent Cloud, along with its tenant, database, and
  API token.
- If you plan to use auto-generated embeddings, an OpenAI-compatible
  embedding endpoint, model, and API key.
- Input records that match the connector’s expected fields: `id`,
  `document`, `embedding`, and `metadata`. If you disable
  auto-generated embeddings, each record must include a pre-computed
  embedding. If you enable auto-generated embeddings, each record must
  include the document text.
- For networking considerations, see
  [Networking and DNS](overview.md#connect-internet-access-resources). To use a set of public egress
  IP addresses, see [Public Egress IP Addresses for Confluent Cloud Connectors](static-egress-ip.md#cc-static-egress-ips).

- Kafka cluster credentials. The following lists the different ways you can provide credentials.
  - Enter an existing [service account](service-account.md#s3-cloud-service-account) resource ID.
  - Create a Confluent Cloud [service account](service-account.md#s3-cloud-service-account) for the connector. Make sure to review the ACL entries required in the [service account documentation](service-account.md#s3-cloud-service-account). Some connectors have specific ACL requirements.
  - Create a Confluent Cloud API key and secret. To create a key and secret, you can use [confluent api-key create](https://docs.confluent.io/confluent-cli/current/command-reference/api-key/confluent_api-key_create.html) *or* you can autogenerate the API key and secret directly in the Cloud Console when setting up the connector.

### Using the Confluent Cloud Console

#### Step 1: Launch your Confluent Cloud cluster

To create and launch a Kafka cluster in Confluent Cloud, see [Create a kafka cluster in Confluent Cloud](../get-started/index.md#cloud-create-kafka-cluster).

#### Step 2: Add a connector

In the left navigation menu, click **Connectors**. If you already have connectors in your cluster, click **+ Add
connector**.

#### Step 3: Select your connector

Click the **ChromaDB Sink** connector card.

![ChromaDB Sink Connector Card](images/cc-chromadb-sink-icon.png)

<a id="cc-chromadb-sink-setup-connection"></a>

#### Step 4: Enter the connector details

Before configuring the connector settings,
ensure that you complete all [prerequisites](#cc-chromadb-sink-prereqs).

#### NOTE
An asterisk ( \* ) designates a required entry.

At the **Add ChromaDB Sink Connector** screen, complete the following:

### Select topic

If you’ve already populated your Kafka topics, select the topics you want
to connect from the **Topics** list.

To create a new topic, click **+Add new topic**.

### Kafka credentials

1. Select the way you want to provide **Kafka Cluster credentials**. You can
   choose one of the following options:
   - **My account**: This setting allows your connector to globally access everything
     that you have access to. With a user account, the connector uses an API key and
     secret to access the Kafka cluster. This option is not recommended for production.
   - **Service account**: This setting limits the access for your connector by using a
     [service account](service-account.md#s3-cloud-service-account). This option is recommended for
     production.
   - **Use an existing API key**: This setting allows you to specify an API key and a
     secret pair. You can use an existing pair or create a new one. This method is not
     recommended for production environments.

   #### NOTE
   Freight clusters support only service accounts for Kafka authentication.
2. Click **Continue**.

### Authentication

1. Configure the authentication properties:

   **ChromaDB Connection**
   - **ChromaDB Endpoint**: The ChromaDB API endpoint URL. For example: `https://api.trychroma.com:8000`.
   - **ChromaDB API Key**: API key for authenticating with ChromaDB Cloud.
   - **Tenant**: The ChromaDB tenant ID. For self-hosted single-tenant deployments, use `default_tenant`. For ChromaDB Cloud, use the tenant ID from the ChromaDB Cloud console.
   - **Database**: The ChromaDB database name. For self-hosted deployments, use `default_database`. For ChromaDB Cloud, use the database name from the ChromaDB Cloud console.
2. Click **Continue**.

### Configuration

#### NOTE
Configuration properties that are not shown in the
Confluent Cloud Console use the default values. For all property values
and definitions, see [Configuration Properties](#cc-chromadb-sink-config-properties).

- **Input Kafka record value format**: Sets the input Kafka record value format. Valid entries are `AVRO`,
  `JSON_SR`, `PROTOBUF`, `JSON`, `BYTES`, or `STRING`. A valid
  schema must be available in [Schema Registry](../get-started/schema-registry.md#cloud-sr-config)
  to use a schema-based message format (for example, Avro, JSON Schema,
  or Protobuf).

**Embedding**

- **Auto-Generate Embeddings**: When enabled, the connector calls an external OpenAI-compatible
  embedding API to generate vector embeddings from the document text.
  When disabled, embeddings must be provided in each Kafka record. Each
  collection declares its own embedding endpoint, model, and API key
  under its own Collection configuration section. There is no shared
  global embedding configuration.

**Collections**

- **Number of collections**: The number of ChromaDB collections to write to. Each collection maps to a set of Kafka topics.
- **Auto-Create Collections**: Whether to automatically create collections if they do not exist.

**Collection 1 configuration**

- **Collection Name**: The name of the ChromaDB collection to write to.
- **Topic**: The Kafka topic to pull data from for this collection.
- **Embedding endpoint**: The full URL of the OpenAI-compatible embedding API for this
  collection (for example, `https://api.openai.com/v1/embeddings`).
  Each collection declares its own embedding endpoint. There is no
  shared global fallback. Required when Auto-Generate Embeddings is
  enabled.
- **Embedding model**: The embedding model for this collection (for example,
  `text-embedding-3-small` or `nomic-embed-text`). Each
  collection’s vectors are dimensioned by its model. ChromaDB rejects
  mixed dimensions within one collection. Required when
  Auto-Generate Embeddings is enabled.
- **Embedding API key**: The API key for the embedding service (for example, OpenAI or Azure OpenAI).

### **Show advanced configurations**

- **Schema context**: Select a schema context to use for this connector, if using
  a schema-based data format. This property defaults to the **Default** context,
  which configures the connector to use the default schema set up for Schema Registry in your
  Confluent Cloud environment. A schema context allows you to use separate schemas (like
  schema sub-registries) tied to topics in different Kafka clusters that share the
  same Schema Registry environment. For example, if you select a non-default context, a
  **Source** connector uses only that schema context to register a schema and a
  **Sink** connector uses only that schema context to read from. For more
  information about setting up a schema context, see [What are schema contexts and when should you use them?](../sr/faqs-cc.md#faq-schema-contexts).

**Additional Configs**

- **Value Converter Schema ID Deserializer**: Sets the class name of the schema ID deserializer for values. The deserializer reads schema IDs from message headers.
- **Value Converter Schemas Enable**: Includes schema within each of the serialized values. Input messages must contain `schema` and `payload` fields and must not contain additional fields. For plain `JSON` data, set this to `false`. Applies to the `JSON` converter.
- **Schema GUID For Key Converter**: Sets the schema GUID to use for deserialization when using `ConfigSchemaIdDeserializer`. This lets you specify a fixed schema GUID for deserializing message keys. This property is applicable only when `key.converter.key.schema.id.deserializer` is set to `ConfigSchemaIdDeserializer`.
- **Value Converter Connect Meta Data**: Enables the Connect converter to add its metadata to the output schema. Applies to Avro converters.
- **Value Converter Value Subject Name Strategy**: Determines how to construct the subject name under which the value schema is registered with Schema Registry.
- **Key Converter Key Subject Name Strategy**: Determines how to construct the subject name for key schema registration.
- **Value Converter Replace Null With Default**: Specifies whether to replace fields that have a default value and that are null to the default value. When set to `true`, the connector uses the default value; otherwise, it uses `null`. Applies to the `JSON` converter.
- **Value Converter Reference Subject Name Strategy**: Sets the subject reference name strategy for values. Valid entries are `DefaultReferenceSubjectNameStrategy` or `QualifiedReferenceSubjectNameStrategy`. You can use this strategy only with `PROTOBUF` format; the default strategy is `DefaultReferenceSubjectNameStrategy`.
- **Schema ID For Value Converter**: Sets the schema ID to use for deserialization when using `ConfigSchemaIdDeserializer`. This lets you specify a fixed schema ID for deserializing message values. This property is applicable only when `value.converter.value.schema.id.deserializer` is set to `ConfigSchemaIdDeserializer`.
- **Errors Tolerance**: Use this property to configure the connector’s error handling behavior.

  #### WARNING
  Use this property with caution for sink connectors, as it can lead to data loss. If you set this property to `all`, the connector does not fail on errant records, but logs them (and sends to DLQ for sink connectors) and continues processing. If you set this property to `none`, the connector task fails on errant records.
- **Value Converter Ignore Default For Nullables**: When set to `true`, this property ensures that the corresponding record in Kafka is `null`, instead of showing the default column value. Applies to the `AVRO`, `PROTOBUF`, and `JSON_SR` converters.
- **Key Converter Schema ID Deserializer**: Sets the class name of the schema ID deserializer for keys. The deserializer reads schema IDs from message headers.
- **Value Converter Decimal Format**: Specifies the `JSON` or `JSON_SR` serialization format for Connect `DECIMAL` logical type values with two allowed literals:
  `BASE64` to serialize `DECIMAL` logical types as base64 encoded binary data, and
  `NUMERIC` to serialize `DECIMAL` logical type values in `JSON` or `JSON_SR` as a number representing the decimal value.
- **Schema GUID For Value Converter**: Sets the schema GUID to use for deserialization when using `ConfigSchemaIdDeserializer`. This lets you specify a fixed schema GUID for deserializing message values. This property is applicable only when `value.converter.value.schema.id.deserializer` is set to `ConfigSchemaIdDeserializer`.
- **Schema ID For Key Converter**: Sets the schema ID to use for deserialization when using `ConfigSchemaIdDeserializer`. This lets you specify a fixed schema ID for deserializing message keys. This property is applicable only when `key.converter.key.schema.id.deserializer` is set to `ConfigSchemaIdDeserializer`.

**Auto-restart policy**

- **Enable Connector Auto-restart**: Enables the auto-restart behavior of the connector and its
  task in the event of user-actionable errors. Defaults to `true`, enabling the connector to
  automatically restart in case of user-actionable errors. Set this property to `false` to
  disable auto-restart for failed connectors. If disabled, you must manually restart the connector.

**Consumer configuration**

- **Max poll interval(ms)**: Sets the maximum delay between subsequent consume requests to Kafka. Use this property to
  improve connector performance in cases when the connector cannot send records to the sink system.
  The default is 300,000 milliseconds (5 minutes).
- **Max poll records**: Sets the maximum number of records to consume from Kafka in a single request. Use this property to
  improve connector performance in cases when the connector cannot send records to the sink system.
  The default is 500 records.

**Collections**

- **Distance Metric**: The distance metric used when the connector creates a collection. Valid values are `cosine` (cosine similarity), `ip` (inner product), and `l2` (squared L2 norm). This setting applies only to collections the connector creates while `chromadb.auto.create.collection` is `true`. A pre-existing collection keeps the space it was created with, and this setting is ignored for it. A single value is shared by every configured collection. Per-collection distance metrics are not supported.

**Collection 1 configuration**

- **Batch Size**: Maximum number of records per upsert request for this collection. Chroma Cloud rejects writes larger than 300 records, so higher values fail with HTTP 422.
- **Behavior on null values**: How to handle Kafka tombstone records (non-null key and null value). `IGNORE` skips the record, and `FAIL` stops the connector.
- **Embedding batch size**: Number of texts to embed per API call for this collection. Higher values improve throughput but use more memory.

**Behavior on error**

- **Behavior On Errors**: Error handling behavior for failed HTTP requests. `IGNORE`
  continues past failed batches. Failed and dropped records are still
  written to the error topic. Set `FAIL` while troubleshooting so the
  task surfaces a mapped error message and stops on the first
  unrecoverable batch.

**Transforms**

- **Single Message Transformations**: To add a new SMT, see [Add transforms](single-message-transforms.md#cc-single-message-transforms-ui).
  For more information about unsupported SMTs, see
  [Unsupported transformations](single-message-transforms.md#cc-single-message-transforms-unsupported-transforms).

**Processing position**

- **Set offsets**: Click **Set offsets** to define a specific offset for
  this connector to begin procession data from. For more information
  on managing offsets, see [Manage offsets](offsets.md#connect-custom-offsets).

For all property values and definitions, see
[Configuration Properties](#cc-chromadb-sink-config-properties).

- Click **Continue**.

### Sizing

Based on the number of topic partitions you select, the connector
recommends a number of tasks.

1. To change the number of recommended tasks, enter the number of
   [tasks](/platform/current/connect/concepts.html#tasks) for the connector to
   use in the **Tasks** field.
2. Click **Continue**.

### Review and Launch

1. Verify the connection details.
2. Click **Launch**.

   The status for the connector should go from **Provisioning** to
   **Running**.

#### Step 5: Check the ChromaDB collection

Verify that new records appear in your ChromaDB collection.

For more information and examples to use with the Confluent Cloud API for Connect,
see the [Confluent Cloud API for Connect Usage Examples](connect-api-section.md#ccloud-connect-api) section.

### Using the Confluent CLI

Complete the following steps to set up and run the connector using the
Confluent CLI.

#### NOTE
Make sure you have all your
[prerequisites](#cc-chromadb-sink-prereqs) completed.

#### Step 1: List the available connectors

Enter the following command to list available connectors:

```none
confluent connect plugin list
```

#### Step 2: List the connector configuration properties

Enter the following command to show the connector configuration properties:

```none
confluent connect plugin describe <connector-plugin-name>
```

The command output shows the required and optional configuration properties.

#### Step 3: Create the connector configuration file

Create a JSON file that contains the connector configuration properties. The
following example shows the required properties for a single collection using
pre-computed embeddings (`chromadb.auto.embed` disabled):

```none
{
  "name": "ChromaDBSink_0",
  "config": {
    "topics": "documents",
    "connector.class": "ChromaDBSink",
    "name": "ChromaDBSink_0",
    "input.data.format": "JSON",
    "kafka.auth.mode": "KAFKA_API_KEY",
    "kafka.api.key": "<my-kafka-api-key>",
    "kafka.api.secret": "<my-kafka-api-secret>",
    "chromadb.endpoint": "https://api.trychroma.com:8000",
    "chromadb.api.key": "<my-chromadb-api-key>",
    "chromadb.tenant": "<my-chromadb-tenant>",
    "chromadb.database": "<my-chromadb-database>",
    "collection1.name": "my-embeddings",
    "collection1.topic": "documents",
    "tasks.max": "1"
  }
}
```

Note the following property definitions:

* `"name"`: Sets a name for your new connector.
* `"connector.class"`: Identifies the connector plugin name.
* `"topics"`: Identifies the topic name or a comma-separated list of
  topic names.
* `"input.data.format"`: Sets the input Kafka record value format.

* `"kafka.auth.mode"`: Identifies the connector authentication mode you want to use. There are two options: `SERVICE_ACCOUNT` or `KAFKA_API_KEY` (the default). To use an API key and secret, specify the configuration properties `kafka.api.key` and `kafka.api.secret`, as shown in the example configuration (above).  To use a [service account](service-account.md#s3-cloud-service-account), specify the **Resource ID** in the property `kafka.service.account.id=<service-account-resource-ID>`. To list the available service account resource IDs, use the following command:
  ```bash
  confluent iam service-account list
  ```

  For example:
  ```bash
  confluent iam service-account list

     Id     | Resource ID |       Name        |    Description
  +---------+-------------+-------------------+-------------------
     123456 | sa-l1r23m   | sa-1              | Service account 1
     789101 | sa-l4d56p   | sa-2              | Service account 2
  ```

* `"chromadb.endpoint"`: The ChromaDB API endpoint URL.
* `"chromadb.api.key"`: The API key for authenticating with Chroma Cloud.
* `"chromadb.tenant"` and `"chromadb.database"`: The ChromaDB tenant
  ID and database name.
* `"collection1.name"`: The name of the ChromaDB collection to write to.
* `"collection1.topic"`: The Kafka topic to pull data from for this collection.

The following are optional properties you can include in the configuration:

* `"collections.num"`: The number of ChromaDB collections to write to.
  Valid values are `1` to `15`. Defaults to `1`.
* `"chromadb.auto.create.collection"`: Whether to automatically create
  collections if they do not exist. Defaults to `true`.
* `"chromadb.distance.metric"`: The distance metric used when the
  connector creates a collection. Valid entries are `cosine` (cosine
  similarity), `ip` (inner product), or `l2` (squared L2 norm).
  Defaults to `cosine`. ChromaDB’s own default space is
  `l2`, so a collection you create yourself can differ from this default.

  This setting applies only to collections the connector creates while
  `chromadb.auto.create.collection` is `true`. A pre-existing
  collection keeps the space it was created with, and this setting is
  ignored for it. Every configured collection shares one value, so
  per-collection distance metrics are not supported.
* `"chromadb.auto.embed"`: Whether the connector generates embeddings by
  calling an OpenAI-compatible embedding API. Defaults to `false`. When
  set to `true`, set `collection1.embedding.endpoint`,
  `collection1.embedding.model`, and `collection1.embedding.api.key`
  for each collection.
* `"collection1.batch.size"`: The maximum number of records per upsert
  request for this collection. Valid values are `1` to `300`.
  Defaults to `50`. ChromaDB Cloud rejects writes larger than 300
  records, so higher values fail with HTTP 422.
* `"collection1.behavior.on.null.values"`: How to handle Kafka tombstone
  records for this collection. Valid entries are `IGNORE` or `FAIL`.
  Defaults to `IGNORE`.
* `"behavior.on.error"`: Error handling behavior for failed HTTP
  requests. Valid entries are `FAIL` or `IGNORE`. Defaults to
  `FAIL`. `IGNORE` continues past failed batches. Failed and dropped
  records are still written to the error topic. Set `FAIL` while
  troubleshooting so the task surfaces a mapped error message and stops
  on the first unrecoverable batch.

**Example: Auto-generated embeddings**

The following example generates embeddings through an OpenAI-compatible API
instead of requiring a pre-computed embedding in each record:

```none
{
  "name": "ChromaDBSink_1",
  "config": {
    "topics": "documents",
    "connector.class": "ChromaDBSink",
    "name": "ChromaDBSink_1",
    "input.data.format": "JSON",
    "kafka.auth.mode": "KAFKA_API_KEY",
    "kafka.api.key": "<my-kafka-api-key>",
    "kafka.api.secret": "<my-kafka-api-secret>",
    "chromadb.endpoint": "https://api.trychroma.com:8000",
    "chromadb.api.key": "<my-chromadb-api-key>",
    "chromadb.tenant": "<my-chromadb-tenant>",
    "chromadb.database": "<my-chromadb-database>",
    "chromadb.auto.embed": "true",
    "collection1.name": "my-embeddings",
    "collection1.topic": "documents",
    "collection1.embedding.endpoint": "https://api.openai.com/v1/embeddings",
    "collection1.embedding.model": "text-embedding-3-small",
    "collection1.embedding.api.key": "<my-openai-api-key>",
    "tasks.max": "1"
  }
}
```

**SMTs**: For details about adding SMTs using the Confluent CLI, see the
[Single Message Transformations](single-message-transforms.md#cc-single-message-transforms)
documentation. For a list of SMTs that are not supported with this connector,
see [Unsupported transformations](single-message-transforms.md#cc-single-message-transforms-unsupported-transforms).

For all property values and
definitions, see [Configuration Properties](#cc-chromadb-sink-config-properties).

#### Step 4: Load the configuration file and create the connector

Enter the following command to load the configuration and start the connector:

```none
confluent connect cluster create --config-file <file-name>.json
```

For example:

```none
confluent connect cluster create --config-file chromadb-sink-config.json
```

Example output:

```none
Created connector ChromaDBSink_0 lcc-ix4dl
```

#### Step 5: Check the connector status

Enter the following command to check the connector status:

```none
confluent connect cluster list
```

Example output:

```none
ID          |        Name        | Status  | Type
+-----------+---------------------+---------+------+
lcc-ix4dl   | ChromaDBSink_0      | RUNNING | sink
```

#### Step 6: Check the ChromaDB collection

Verify that new records appear in your ChromaDB collection.

For more information and examples to use with the Confluent Cloud API for Connect,
see the [Confluent Cloud API for Connect Usage Examples](connect-api-section.md#ccloud-connect-api) section.

<a id="cc-chromadb-sink-config-properties"></a>

## Configuration Properties

Use the following configuration properties with the fully managed connector.

### Which topics do you want to get data from?

`topics.regex`
: A regular expression that matches the names of the topics to consume from. This is useful when you want to consume from multiple topics that match a certain pattern without having to list them all individually.
  <br/>
  * Type: string
  * Importance: low

`topics`
: Identifies the topic name or a comma-separated list of topic names.
  <br/>
  * Type: list
  * Importance: high

`errors.deadletterqueue.topic.name`
: The name of the topic to be used as the dead letter queue (DLQ) for messages that result in an error when processed by this sink connector, or its transformations or converters. Defaults to ‘dlq-${connector}’ if not set. The DLQ topic will be created automatically if it does not exist. You can provide `${connector}` in the value to use it as a placeholder for the logical cluster ID.
  <br/>
  * Type: string
  * Default: dlq-${connector}
  * Importance: low

`reporter.result.topic.name`
: The name of the topic to produce records to after successfully processing a sink record. Defaults to ‘success-${connector}’ if not set. You can provide `${connector}` in the value to use it as a placeholder for the logical cluster ID.
  <br/>
  * Type: string
  * Default: success-${connector}
  * Importance: low

`reporter.error.topic.name`
: The name of the topic to produce records to after each unsuccessful record sink attempt. Defaults to ‘error-${connector}’ if not set. You can provide `${connector}` in the value to use it as a placeholder for the logical cluster ID.
  <br/>
  * Type: string
  * Default: error-${connector}
  * Importance: low

### Schema Config

`schema.context.name`
: Add a schema context name. A schema context represents an independent scope in Schema Registry. It is a separate sub-schema tied to topics in different Kafka clusters that share the same Schema Registry instance. If not used, the connector uses the default schema configured for Schema Registry in your Confluent Cloud environment.
  <br/>
  * Type: string
  * Default: default
  * Importance: medium

### Input messages

`input.data.format`
: Sets the input Kafka record value format. Valid entries are AVRO, JSON_SR, PROTOBUF, JSON, BYTES or STRING. Note that you need to have Confluent Cloud Schema Registry configured if using a schema-based message format like AVRO, JSON_SR, and PROTOBUF.
  <br/>
  * Type: string
  * Default: JSON
  * Importance: high

### How should we connect to your data?

`name`
: Sets a name for your connector.
  <br/>
  * Type: string
  * Valid Values: A string at most 64 characters long
  * Importance: high

### Kafka Cluster credentials

`kafka.auth.mode`
: Kafka Authentication mode. It can be one of KAFKA_API_KEY or SERVICE_ACCOUNT. It defaults to KAFKA_API_KEY mode, whenever possible.
  <br/>
  * Type: string
  * Valid Values: SERVICE_ACCOUNT, KAFKA_API_KEY
  * Importance: high

`kafka.api.key`
: Kafka API Key. Required when kafka.auth.mode==KAFKA_API_KEY.
  <br/>
  * Type: password
  * Importance: high

`kafka.service.account.id`
: The Service Account that will be used to generate the API keys to communicate with Kafka Cluster.
  <br/>
  * Type: string
  * Importance: high

`kafka.api.secret`
: Secret associated with Kafka API key. Required when kafka.auth.mode==KAFKA_API_KEY.
  <br/>
  * Type: password
  * Importance: high

### Consumer configuration

`max.poll.interval.ms`
: The maximum delay between subsequent consume requests to Kafka. This configuration property may be used to improve the performance of the connector, if the connector cannot send records to the sink system. Defaults to 300000 milliseconds (5 minutes).
  <br/>
  * Type: long
  * Default: 300000 (5 minutes)
  * Valid Values: [60000,…,1800000] for non-dedicated clusters and [60000,…] for dedicated clusters
  * Importance: low

`max.poll.records`
: The maximum number of records to consume from Kafka in a single request. This configuration property may be used to improve the performance of the connector, if the connector cannot send records to the sink system. Defaults to 500 records.
  <br/>
  * Type: long
  * Default: 500
  * Valid Values: [1,…,500] for non-dedicated clusters and [1,…] for dedicated clusters
  * Importance: low

### Number of tasks for this connector

`tasks.max`
: Maximum number of tasks for the connector.
  <br/>
  * Type: int
  * Valid Values: [1,…]
  * Importance: high

### ChromaDB Connection

`chromadb.endpoint`
: The ChromaDB API endpoint URL. For example: `https://api.trychroma.com:8000`.
  <br/>
  * Type: string
  * Importance: high

`chromadb.api.key`
: API key for authenticating with ChromaDB Cloud.
  <br/>
  * Type: password
  * Importance: high

`chromadb.tenant`
: The ChromaDB tenant ID. For self-hosted single-tenant deployments, use `default_tenant`. For ChromaDB Cloud, use the tenant ID from the Cloud console.
  <br/>
  * Type: string
  * Importance: high

`chromadb.database`
: The ChromaDB database name. For self-hosted deployments, use `default_database`. For ChromaDB Cloud, use the database name from the Cloud console.
  <br/>
  * Type: string
  * Importance: high

### Embedding

`chromadb.auto.embed`
: When enabled, the connector calls an external OpenAI-compatible embedding API to generate vector embeddings from the document text. When disabled, embeddings must be provided in each Kafka record. Each collection declares its own embedding endpoint, model, and API key under its Collection {i} configuration section. There is no shared global embedding configuration.
  <br/>
  * Type: boolean
  * Default: false
  * Importance: high

### Collections

`collections.num`
: The number of ChromaDB collections to write to. Each collection maps to a set of Kafka topics. Valid values are `1` to `15`. Defaults to `1`.
  <br/>
  * Type: int
  * Default: 1
  * Valid Values: [1,…,15]
  * Importance: high

`chromadb.auto.create.collection`
: Whether to automatically create collections if they do not exist.
  <br/>
  * Type: boolean
  * Default: true
  * Importance: medium

`chromadb.distance.metric`
: The distance metric for newly created collections.
  <br/>
  * Type: string
  * Default: cosine
  * Valid Values: cosine, ip, l2
  * Importance: medium

### Collection 1 configuration

`collection1.name`
: The name of the ChromaDB collection to write to.
  <br/>
  * Type: string
  * Importance: high

`collection1.topic`
: Kafka topic to pull data from for this collection.
  <br/>
  * Type: string
  * Default: “”
  * Importance: high

`collection1.batch.size`
: Maximum number of records per upsert request for this collection.
  <br/>
  * Type: int
  * Default: 50
  * Valid Values: [1,…,500]
  * Importance: medium

`collection1.behavior.on.null.values`
: How to handle Kafka tombstone records (non-null key and null value). `IGNORE` skips the record. `FAIL` stops the connector.
  <br/>
  * Type: string
  * Default: IGNORE
  * Valid Values: FAIL, IGNORE
  * Importance: low

`collection1.embedding.endpoint`
: Full URL of the OpenAI-compatible embedding API for this collection (for example, `https://api.openai.com/v1/embeddings`). Each collection declares its own. There is no shared global fallback. This field is required when Auto-Generate Embeddings is enabled.
  <br/>
  * Type: string
  * Importance: high

`collection1.embedding.model`
: Embedding model for this collection (for example, `text-embedding-3-small`, `nomic-embed-text`). Each collection’s vectors are dimensioned by its model. Mixing dimensions in one collection would be rejected by ChromaDB. This field is required when Auto-Generate Embeddings is enabled.
  <br/>
  * Type: string
  * Importance: high

`collection1.embedding.api.key`
: API key for the embedding service (for example, OpenAI or Azure OpenAI).
  <br/>
  * Type: password
  * Importance: high

`collection1.embedding.batch.size`
: Number of texts to embed per API call for this collection. Higher values improve throughput but use more memory.
  <br/>
  * Type: int
  * Default: 10
  * Valid Values: [1,…,100]
  * Importance: low

### Behavior on error

`behavior.on.error`
: Error handling behavior for failed HTTP requests. `FAIL` stops the connector on the first error after retries are exhausted. `IGNORE` reports the failed batch to the error topic and continues with the next batch. Both modes route batches to the configured error topic through the reporter. `IGNORE` does not silently drop records.
  <br/>
  * Type: string
  * Default: FAIL
  * Valid Values: FAIL, IGNORE
  * Importance: low

### Additional Configs

`consumer.override.auto.offset.reset`
: Defines the behavior of the consumer when there is no committed position (which occurs when the group is first initialized) or when an offset is out of range. You can choose either to reset the position to the “earliest” offset (the default) or the “latest” offset. You can also select “none” if you would rather set the initial offset yourself and you are willing to handle out of range errors manually. More details: [https://docs.confluent.io/platform/current/installation/configuration/consumer-configs.html#auto-offset-reset](https://docs.confluent.io/platform/current/installation/configuration/consumer-configs.html#auto-offset-reset)
  <br/>
  * Type: string
  * Importance: low

`consumer.override.isolation.level`
: Controls how to read messages written transactionally. If set to read_committed, consumer.poll() will only return transactional messages which have been committed. If set to read_uncommitted (the default), consumer.poll() will return all messages, even transactional messages which have been aborted. Non-transactional messages will be returned unconditionally in either mode.  More details: [https://docs.confluent.io/platform/current/installation/configuration/consumer-configs.html#isolation-level](https://docs.confluent.io/platform/current/installation/configuration/consumer-configs.html#isolation-level)
  <br/>
  * Type: string
  * Importance: low

`header.converter`
: The converter class for the headers. This is used to serialize and deserialize the headers of the messages.
  <br/>
  * Type: string
  * Importance: low

`key.converter.use.schema.guid`
: The schema GUID to use for deserialization when using ConfigSchemaIdDeserializer. This allows you to specify a fixed schema GUID to be used for deserializing message keys. Only applicable when key.converter.key.schema.id.deserializer is set to ConfigSchemaIdDeserializer.
  <br/>
  * Type: string
  * Importance: low

`key.converter.use.schema.id`
: The schema ID to use for deserialization when using ConfigSchemaIdDeserializer. This allows you to specify a fixed schema ID to be used for deserializing message keys. Only applicable when key.converter.key.schema.id.deserializer is set to ConfigSchemaIdDeserializer.
  <br/>
  * Type: int
  * Importance: low

`value.converter.allow.optional.map.keys`
: Allow optional string map key when converting from Connect Schema to Avro Schema. Applicable for Avro Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.auto.register.schemas`
: Specify if the Serializer should attempt to register the Schema.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.connect.meta.data`
: Allow the Connect converter to add its metadata to the output schema. Applicable for Avro Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.enhanced.avro.schema.support`
: Enable enhanced schema support to preserve package information and Enums. Applicable for Avro Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.enhanced.protobuf.schema.support`
: Enable enhanced schema support to preserve package information. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.flatten.unions`
: Whether to flatten unions (oneofs). Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.generate.index.for.unions`
: Whether to generate an index suffix for unions. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.generate.struct.for.nulls`
: Whether to generate a struct variable for null values. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.int.for.enums`
: Whether to represent enums as integers. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.latest.compatibility.strict`
: Verify latest subject version is backward compatible when use.latest.version is true.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.object.additional.properties`
: Whether to allow additional properties for object schemas. Applicable for JSON_SR Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.optional.for.nullables`
: Whether nullable fields should be specified with an optional label. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.optional.for.proto2`
: Whether proto2 optionals are supported. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.scrub.invalid.names`
: Whether to scrub invalid names by replacing invalid characters with valid characters. Applicable for Avro and Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.use.latest.version`
: Use latest version of schema in subject for serialization when auto.register.schemas is false.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.use.optional.for.nonrequired`
: Whether to set non-required properties to be optional. Applicable for JSON_SR Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.use.schema.guid`
: The schema GUID to use for deserialization when using ConfigSchemaIdDeserializer. This allows you to specify a fixed schema GUID to be used for deserializing message values. Only applicable when value.converter.value.schema.id.deserializer is set to ConfigSchemaIdDeserializer.
  <br/>
  * Type: string
  * Importance: low

`value.converter.use.schema.id`
: The schema ID to use for deserialization when using ConfigSchemaIdDeserializer. This allows you to specify a fixed schema ID to be used for deserializing message values. Only applicable when value.converter.value.schema.id.deserializer is set to ConfigSchemaIdDeserializer.
  <br/>
  * Type: int
  * Importance: low

`value.converter.wrapper.for.nullables`
: Whether nullable fields should use primitive wrapper messages. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`value.converter.wrapper.for.raw.primitives`
: Whether a wrapper message should be interpreted as a raw primitive at root level. Applicable for Protobuf Converters.
  <br/>
  * Type: boolean
  * Importance: low

`errors.tolerance`
: Use this property if you would like to configure the connector’s error handling behavior. WARNING: This property should be used with CAUTION for SOURCE CONNECTORS as it may lead to dataloss. If you set this property to ‘all’, the connector will not fail on errant records, but will instead log them (and send to DLQ for Sink Connectors) and continue processing. If you set this property to ‘none’, the connector task will fail on errant records.
  <br/>
  * Type: string
  * Default: all
  * Importance: low

`key.converter.key.schema.id.deserializer`
: The class name of the schema ID deserializer for keys. This is used to deserialize schema IDs from the message headers.
  <br/>
  * Type: string
  * Default: io.confluent.kafka.serializers.schema.id.DualSchemaIdDeserializer
  * Importance: low

`key.converter.key.subject.name.strategy`
: How to construct the subject name for key schema registration.
  <br/>
  * Type: string
  * Default: TopicNameStrategy
  * Importance: low

`value.converter.decimal.format`
: Specify the JSON/JSON_SR serialization format for Connect DECIMAL logical type values with two allowed literals:
  <br/>
  BASE64 to serialize DECIMAL logical types as base64 encoded binary data and
  <br/>
  NUMERIC to serialize Connect DECIMAL logical type values in JSON/JSON_SR as a number representing the decimal value.
  <br/>
  * Type: string
  * Default: BASE64
  * Importance: low

`value.converter.flatten.singleton.unions`
: Whether to flatten singleton unions. Applicable for Avro and JSON_SR Converters.
  <br/>
  * Type: boolean
  * Default: false
  * Importance: low

`value.converter.ignore.default.for.nullables`
: When set to true, this property ensures that the corresponding record in Kafka is NULL, instead of showing the default column value. Applicable for AVRO,PROTOBUF and JSON_SR Converters.
  <br/>
  * Type: boolean
  * Default: false
  * Importance: low

`value.converter.reference.subject.name.strategy`
: Set the subject reference name strategy for value. Valid entries are DefaultReferenceSubjectNameStrategy or QualifiedReferenceSubjectNameStrategy. Note that the subject reference name strategy can be selected only for PROTOBUF format with the default strategy being DefaultReferenceSubjectNameStrategy.
  <br/>
  * Type: string
  * Default: DefaultReferenceSubjectNameStrategy
  * Importance: low

`value.converter.replace.null.with.default`
: Whether to replace fields that have a default value and that are null to the default value. When set to true, the default value is used, otherwise null is used. Applicable for JSON Converter.
  <br/>
  * Type: boolean
  * Default: true
  * Importance: low

`value.converter.schemas.enable`
: Include schemas within each of the serialized values. Input messages must contain schema and payload fields and may not contain additional fields. For plain JSON data, set this to false. Applicable for JSON Converter.
  <br/>
  * Type: boolean
  * Default: false
  * Importance: low

`value.converter.value.schema.id.deserializer`
: The class name of the schema ID deserializer for values. This is used to deserialize schema IDs from the message headers.
  <br/>
  * Type: string
  * Default: io.confluent.kafka.serializers.schema.id.DualSchemaIdDeserializer
  * Importance: low

`value.converter.value.subject.name.strategy`
: Determines how to construct the subject name under which the value schema is registered with Schema Registry.
  <br/>
  * Type: string
  * Default: TopicNameStrategy
  * Importance: low

### Auto-restart policy

`auto.restart.on.user.error`
: Enable connector to automatically restart on user-actionable errors.
  <br/>
  * Type: boolean
  * Default: true
  * Importance: medium

## Frequently asked questions

Find answers to frequently asked questions about the ChromaDB Sink
connector for Confluent Cloud.

### Why does the connector fail with an HTTP 422 error?

An HTTP 422 response from ChromaDB usually means the upsert batch is too
large or the records in the batch don’t have consistent embedding
dimensions.

**Common causes:**

* **Batch size too large for ChromaDB Cloud**: ChromaDB Cloud enforces a
  per-request upsert limit of approximately 300 records per request. If you
  increase `collection1.batch.size` above this limit, ChromaDB rejects
  the request.
* **Embedding dimension mismatch**: All embeddings in a batch, and in the
  target collection, must share the same dimension. A record with a
  different-sized embedding than the rest of the collection triggers
  this error.

**Resolution:**

1. Set `collection1.batch.size` (and the equivalent property for any
   other configured collection) to `250` or lower when writing to
   ChromaDB Cloud.
2. Confirm that every record’s `embedding` field has the same number
   of dimensions as the target collection, and that the embedding model
   generating them hasn’t changed.

### Why do I see errors about mixed batches or inconsistent embeddings?

This error occurs when `chromadb.auto.embed` is disabled and a batch
contains some records with a populated `embedding` field and some
without. ChromaDB requires that a single upsert either includes
embeddings for every record or for none of them. The same applies to
dimensions: every embedding in one upsert must be the same length.

What the connector does with such a batch depends on
`behavior.on.error`:

* `FAIL` (the default): the connector rejects the whole batch, writes
  it to the error topic, and stops the task.
* `IGNORE`: the connector removes only the inconsistent records, writes
  those to the error topic with a reason of
  `Record dropped before ChromaDB upsert`, and upserts the rest.

In both modes the affected records are written to the error topic. They
are never discarded without a record of the failure.

**Resolution:**

1. Ensure every record sent to the connector includes a pre-computed
   `embedding` field, or
2. Set `chromadb.auto.embed` to `true` so the connector generates
   embeddings for records that are missing one.

### Why is the connector failing with an authentication error?

A `401` response can come from either ChromaDB itself or, in
auto-embed mode, from the embedding provider.

**Common causes:**

* **Incorrect or rotated** `chromadb.api.key`: the key configured on
  the connector no longer matches the key on your ChromaDB target.
* **Incorrect or expired embedding API key**: when
  `chromadb.auto.embed` is `true`, an invalid
  `collection1.embedding.api.key` causes the embedding call itself to
  fail with a 401.

**Resolution:**

1. Verify `chromadb.api.key` matches the current key on your ChromaDB
   Cloud account or self-hosted deployment.
2. If you use auto-generated embeddings, verify
   `collection1.embedding.api.key` is valid for your embedding
   provider.
3. After updating credentials, restart the connector to apply the
   change.

### Why do I see warnings that a ChromaDB collection wasn’t found?

When `chromadb.auto.create.collection` is set to `true` (the
default), a brief “collection not found” warning is expected the first
time the connector targets a new or recently deleted collection. The
connector automatically recreates the collection and retries, typically
within a few seconds, with no data loss.

This becomes a persistent issue only when:

* `chromadb.auto.create.collection` is set to `false` and the
  target collection doesn’t exist, or
* the configured `chromadb.api.key` doesn’t have permission to create
  collections in your ChromaDB tenant.

**Resolution:** Set `chromadb.auto.create.collection` to `true`, or
manually create the collection and verify the API key’s permissions.

### How can I improve the connector’s write throughput?

Throughput to a single ChromaDB collection scales with `tasks.max` up
to a point, then can plateau or decline. This happens because ChromaDB
serializes writes to a given collection on its side, not because of a
connector limitation.

**Resolution:**

1. Increase `tasks.max` to scale throughput against a single
   collection, but avoid over-provisioning tasks beyond what one
   collection can absorb.
2. To scale further, distribute records across multiple collections
   using `collections.num` and multiple `collection{n}.topic`
   mappings, rather than continuing to raise `tasks.max` on one
   collection.
3. Deploy the connector in the same region as your ChromaDB target to
   avoid cross-region latency on every write.

### Why is throughput lower when auto-generated embeddings are enabled?

With `chromadb.auto.embed` set to `true`, the connector calls an
external embedding API for every batch. End-to-end throughput is then
bound by that embedding provider’s latency and rate limits, not by the
connector.

**Resolution:** For the highest sustained throughput, generate
embeddings upstream, for example with a stream processing job, and
supply them in the `embedding` field with `chromadb.auto.embed` set
to `false`. This decouples sink throughput from embedding API
latency.

### The connector is running but no data appears. What should I check?

**Common causes:**

* **Empty source topic**: verify the source topic actually has
  unconsumed records. A connector with nothing to read appears idle
  rather than failed.
* **Records dropped during preprocessing**: with `chromadb.auto.embed`
  set to `false` and `behavior.on.error` set to `IGNORE`, records
  with a null or empty `embedding` field are removed from the upsert.
  They are written to the error topic rather than discarded, so check
  the error topic before assuming the records were lost.
* `behavior.on.error` **set to** `IGNORE`: the connector continues
  past failed batches, so the task stays in a `Running` state while
  individual batches fail. Set `behavior.on.error` to `FAIL` while
  troubleshooting so the task surfaces a mapped error message and stops
  on the first unrecoverable batch.

**Resolution:**

1. Confirm new records are actually being produced to the source topic.
2. Check that every record includes a valid, non-empty `embedding`
   field, or enable `chromadb.auto.embed`.
3. Set `behavior.on.error` to `FAIL` and review the connector logs
   or dead letter queue for rejected records.

## Next Steps

For an example that shows fully managed Confluent Cloud connectors in action with
Confluent Cloud for Apache Flink, see the [Cloud ETL Demo](/platform/current/tutorials/examples/cloud-etl/docs/index.html).
This example also shows how to use Confluent CLI to manage your resources in
Confluent Cloud.

[![image](images/topology.png)](https://docs.confluent.io/platform/current/tutorials/examples/cloud-etl/docs/index.html)