Tracker Stores

• InMemoryTrackerStore: Stores conversations in memory (default). Suitable for development and testing.
• SQLTrackerStore: Uses SQL databases (PostgreSQL, SQLite, Oracle) for persistent storage.
• RedisTrackerStore: Leverages Redis for fast in-memory storage with optional persistence.
• MongoTrackerStore: Stores conversations in MongoDB, a document-oriented NoSQL database.
• DynamoTrackerStore: Uses AWS DynamoDB for cloud-based storage.

​

User-Scoped Tracker Querying

​

Conversation Properties

• user_id: The unique identifier for the user associated with the conversation. This field is used for querying all conversations belonging to a specific user. Serialization of trackers omits null user_id values to optimize storage.
• conversation_started_timestamp: The timestamp when the conversation was first started. This field is used for ordering conversations chronologically.
• current_session_id: The session ID for the tracker’s current session, derived from the metadata of the last stored event. This value is null when the last event is ConversationInactive, correctly reflecting that no active session exists.

​

InMemoryTrackerStore (default)

​

Configuration

​

User-Scoped Querying

​

Implementation Details

• Full Scan: The implementation scans all stored tracker keys in memory and retrieves each tracker to filter by user_id. This approach is feasible for in-memory storage due to the typically smaller dataset size.
• Ordering and Pagination: After filtering, the trackers are ordered by conversation_started_timestamp and sender_id, and paginated in memory using standard Python sorting and slicing techniques.

​

SQLTrackerStore

​

Configuration

1. Add required configuration to your endpoints.yml:

tracker_store:
    type: SQL
    dialect: "postgresql"  # the dialect used to interact with the db
    url: ""  # (optional) host of the sql db, e.g. "localhost"
    db: "rasa"  # path to your db
    username:  # username used for authentication
    password:  # password used for authentication
    query: # optional dictionary to be added as a query string to the connection URL
      driver: my-driver

2. To start the Rasa server using your SQL backend, add the --endpoints flag, e.g.:

rasa run -m models --endpoints endpoints.yml

​

Configuration Parameters

• domain (default: None): Domain object associated with this tracker store
• dialect (default: sqlite): The dialect used to communicate with your SQL backend. Consult the SQLAlchemy docs for available dialects.
• url (default: None): URL of your SQL server
• port (default: None): Port of your SQL server
• db (default: rasa.db): The path to the database to be used
• username (default: None): The username which is used for authentication
• password (default: None): The password which is used for authentication
• event_broker (default: None): Event broker to publish events to
• login_db (default: None): Alternative database name to which initially connect, and create the database specified by db (PostgreSQL only)
• query (default: None): Dictionary of options to be passed to the dialect and/or the DBAPI upon connect

​

Compatible Databases

• PostgreSQL
• Oracle > 11.0
• SQLite

​

Configuring Oracle

CREATE SEQUENCE username.events_seq;

FROM rasa/rasa:latest-full

# Switch to root user to install packages
USER root

RUN apt-get update -qq && apt-get install -y --no-install-recommends alien libaio1 && apt-get clean && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*

# Copy in oracle instaclient
# https://www.oracle.com/database/technologies/instant-client/linux-x86-64-downloads.html
COPY oracle.rpm oracle.rpm

# Install the Python wrapper library for the Oracle drivers
RUN pip install cx-Oracle

# Install Oracle client libraries
RUN alien -i oracle.rpm

USER 1001

docker build . -t rasa-oracle:latest-oracle-full

​

Using IAM roles to authenticate to an AWS RDS SQL tracker store

GRANT rds_iam TO <db_username>;

{    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "rds-db:connect"
            ],
            "Resource": [
                "arn:aws:rds-db:<region>:<account-id>:dbuser:<DbiResourceId for RDS PostgreSQL>/<db_username>"
            ]
        }
    ]
}

• IAM_CLOUD_PROVIDER: Set this to aws.
• AWS_DEFAULT_REGION: Set this to the AWS region where your RDS database is located.
• RDS_SQL_DB_AWS_IAM_ENABLED: Set this to true to enable IAM authentication for RDS connections.
• SQL_TRACKER_STORE_SSL_MODE: Set this to the desired SSL mode for the connection (Check your deployed SQL database documentation for supported values. For example: see PostgreSQL SSL Mode descriptions).
• SQL_TRACKER_STORE_SSL_ROOT_CERTIFICATE: (Optional) Set this to the path of the CA certificate to use for SSL verification if using verify-ca or verify-full SSL modes. This can be downloaded for the particular region from here.

tracker_store:
    type: sql
    dialect: "postgresql"
    url: "<rds-endpoint>"
    db: "<db-name>"
    username: "<db-username>"
    port: 5432

​

User-Scoped Querying

​

Implementation Details

• Users Table: A separate users table is created with columns for sender_id (mapped to user_id) and conversation_started_timestamp. The table uses dialect-specific upsert operations for PostgreSQL and sqlite and a generic approach (update followed by insert if needed) for other SQL databases to maintain compatibility.
• Bulk Retrieval: User-scoped queries use two bulk operations: a paginated users table query to retrieve sender_ids for the given user_id, followed by a single events WHERE sender_id IN (...) fetch to load all events for those conversations. This ensures query count stays constant regardless of conversation volume.
• Database-Level Ordering and Pagination: The sender_ids are ordered by conversation_started_timestamp and paginated using SQL LIMIT/OFFSET clauses for optimal performance.
• Automatic Cleanup: When a conversation is deleted, the corresponding entry in the users table is automatically cleaned up to maintain referential integrity.
• Efficient Indexing: Appropriate indices are created on the users table for all fields to optimize query performance for user-scoped lookups.

​

RedisTrackerStore

​

High Availability Support

• Redis Cluster Mode: Provides horizontal scaling and is compatible with cloud-hosted Redis services.
• Redis Sentinel Mode: Offers high availability through automatic master/slave failover.
• Standard Mode: Maintains backward compatibility with existing single-instance deployments.

​

Configuration

1. Add required configuration to your endpoints.yml:

tracker_store:
    type: redis
    url: <url of the redis instance, e.g. localhost>
    port: <port of your redis instance, usually 6379>
    key_prefix: <alphanumeric value to prepend to tracker store keys>
    db: <number of your database within redis, e.g. 0. Not used in cluster mode>
    password: <password used for authentication>
    use_ssl: <whether or not the communication is encrypted, default `false`>
    # high availability parameters introduced in 3.14
    deployment_mode: <standard, cluster, or sentinel>
    endpoints: <list of redis cluster/sentinel node addresses in the format host:port, only used in cluster or sentinel mode>
    sentinel_service: <name of the redis sentinel service, only used in sentinel mode>

2. To start the Rasa server using your SQL backend, add the --endpoints flag, e.g.:

rasa run -m models --endpoints endpoints.yml

​

Configuration Parameters

• url (default: localhost): The url of your redis instance
• port (default: 6379): The port which redis is running on
• db (default: 0): The number of your redis database. Ignored in cluster mode as Redis Cluster always uses database 0
• key_prefix (default: None): The prefix to prepend to tracker store keys. Must be alphanumeric
• username (default: None): Username used for authentication
• password (default: None): Password used for authentication (None equals no authentication)
• record_exp (default: None): Record expiry in seconds
• use_ssl (default: False): whether or not to use SSL for transit encryption
• deployment_mode (default: standard): Deployment mode of Redis. One of standard, cluster, or sentinel.
• endpoints (default: None): List of Redis cluster node addresses in the format host:port. Used for cluster and sentinel modes. For cluster mode, these are cluster node endpoints. For sentinel mode, these are sentinel instance endpoints.
• sentinel_service (default: mymaster): Name of the Redis sentinel service. Only used in sentinel mode.

Deployment Mode Details

• Standard Mode: Connects to a single Redis instance using url and port parameters.
• Cluster Mode: Connects to a Redis Cluster for horizontal scaling. If endpoints is not provided, falls back to auto-discovery using url and port.
• Sentinel Mode: Connects to Redis Sentinel for high availability with automatic master/slave failover.

​

Choosing a Deployment Mode

• Use standard mode for simple deployments with a single Redis instance.
• Use cluster mode for horizontal scaling and when using cloud Redis services that require cluster mode.
• Use sentinel mode for high availability with master/slave replication and automatic failover.

​

Using IAM to authenticate to AWS ElastiCache for Redis

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "elasticache:Connect"
            ],
            "Resource": "*"
        }
    ]
}

• IAM_CLOUD_PROVIDER: Set this to aws.
• AWS_DEFAULT_REGION: Set this to the AWS region where your ElastiCache cluster or replication group is located.
• AWS_ELASTICACHE_CLUSTER_NAME: Set this to the name of your ElastiCache cluster or replication group.
• ELASTICACHE_REDIS_AWS_IAM_ENABLED: Set this to true to enable IAM authentication for ElastiCache connections.

tracker_store:
    type: redis
    url: "master redis host of your elasticache cluster or replication group"
    port: "6379"
    use_ssl: true
    username: your-iam-username

​

User-Scoped Querying

​

Implementation Details

• Secondary Index: A Redis Sorted Set is maintained with the key pattern user_trackers:{user_id}. This index stores all conversation IDs (sender_ids) for each user, enabling efficient O(1) lookup.
• Batch Operations: When retrieving multiple conversations for a user, the tracker store uses Redis MGET (multi-get) operations to fetch all tracker data in a single batch, reducing network round trips.
• Index Cleanup: When a conversation is deleted, the corresponding entry is removed from the user’s sorted-set index. Orphaned index entries (entries whose tracker key no longer exists in Redis) are pruned in a single batched ZREM call rather than one call per orphan.

​

MongoTrackerStore

​

Configuration

1. Add required configuration to your endpoints.yml:

tracker_store:
    type: mongod
    url: <url to your mongo instance, e.g. mongodb://localhost:27017>
    db: <name of the db within your mongo instance, e.g. rasa>
    username: <username used for authentication>
    password: <password used for authentication>
    auth_source: <database name associated with the user's credentials>

2. To start the Rasa server using your configured MongoDB instance, add the --endpoints flag, for example:

rasa run -m models --endpoints endpoints.yml

​

Configuration Parameters

• url (default: mongodb://localhost:27017): URL of your MongoDB
• db (default: rasa): The database name which should be used
• username (default: 0): The username which is used for authentication
• password (default: None): The password which is used for authentication
• auth_source (default: admin): database name associated with the user’s credentials.
• collection (default: conversations): The collection name which is used to store the conversations

​

User-Scoped Querying

​

Implementation Details

• Database Indices: The tracker store creates indices on user_id and a compound index on (conversation_started_timestamp, sender_id) to optimize query performance.
• Aggregation Pipeline: User-scoped queries use MongoDB aggregation pipelines to efficiently filter, sort, and paginate conversations. This allows for complex queries with ordering and pagination at the database level.
• User ID Restoration: The implementation ensures that user_id is properly restored from stored tracker data, maintaining data consistency across conversation sessions.

​

DynamoTrackerStore

​

Configuration

1. Add required configuration to your endpoints.yml:

tracker_store:
    type: dynamo
    table_name: <name of the table to create, e.g. rasa>
    region: <name of the region associated with the client>

2. To start the Rasa server using your configured DynamoDB instance, add the --endpoints flag, e.g.:

rasa run -m models --endpoints endpoints.yml

​

Configuration Parameters

• table_name (default: states): name of the DynamoDB table
• region (default: us-east-1): name of the region associated with the client

​

User-Scoped Querying

​

Implementation Details

• Metadata Persistence: The tracker store saves both user_id and conversation_started_timestamp (stored as Decimal type) in the update_item db operation when saving trackers.
• Global Secondary Index (GSI): A GSI named user_id-index is created with user_id as the partition key and conversation_started_timestamp as the sort key.
• Full-Scan Fallback: If the GSI is not available or not yet created, the implementation falls back to a full table scan which uses a filter expression by user_id.

​

Custom Tracker Store

super().__init__(domain, event_broker, **kwargs)

• save: saves the conversation to the tracker store. Must respect the following signature:

    async def save(self, tracker: DialogueStateTracker) -> None:
        """Save tracker to the tracker store.

        Args:
            tracker: Tracker to be saved.
        """

• update: updates an existing tracker in the store (e.g. after anonymization or when retaining events during deletion). When supported, this allows a single atomic write instead of delete-then-save. The PII anonymization and deletion jobs use this method (deletion only when events need to be retained). Must respect the following signature:

    async def update(
        self,
        tracker_to_keep: DialogueStateTracker,
        apply_deletion_only: bool = True,
    ) -> None:
        """Update existing tracker in the tracker store.

        Args:
            tracker_to_keep: Tracker with updated content (e.g. anonymized events or retained events after deletion).
            apply_deletion_only: If True (default), only apply deletion semantics; if False, perform a full content update (e.g. for anonymization).
        """

• retrieve: retrieves tracker for the latest conversation session. Must respect the following signature:

    async def retrieve(self, sender_id: str) -> Optional[DialogueStateTracker]:
        """Retrieve tracker for the given sender_id.

        Args:
            sender_id: Conversation ID to retrieve the tracker for.

        Returns:
            Tracker for the given sender_id.
        """

• keys: returns the set of values for the tracker store’s primary key. Must respect the following signature:

    async def keys(self) -> Iterable[str]:
        """Return the set of values for the tracker store's primary key."""

• delete: deletes the conversation corresponding to the given sender_id in the tracker store. Must respect the following signature:

    async def delete(self, sender_id: str) -> None:
        """Delete tracker for the given sender_id.

        Args:
            sender_id: Conversation ID to delete the tracker for.
        """

• get_serialized_trackers_by_user_id: retrieves serialized event dicts for all conversations belonging to a specific user, with pagination and ordering support. This method powers the GET /users/{user_id}/trackers endpoint and returns data directly from storage without replaying events through DialogueStateTracker. Must respect the following signature:

    async def get_serialized_trackers_by_user_id(
        self,
        user_id: str,
        limit: Optional[int] = None,
        skip: Optional[int] = None,
    ) -> List[Dict[str, Any]]:
        """Retrieve serialized trackers for the given user_id.

        Args:
            user_id: User ID to retrieve trackers for.
            limit: Optional limit on the number of trackers to return.
            skip: Optional offset for pagination.

        Returns:
            List of serialized tracker dicts for the given user_id, ordered
            by conversation_started_timestamp and sender_id, without replaying
            events through DialogueStateTracker.
        """

• _sort_key_serialized(self, tracker_dict: Dict[str, Any]) -> Tuple[float, str]: Returns a tuple of (conversation_started_timestamp, sender_id) for sorting trackers.
• _apply_pagination_serialized(self, trackers: List[Dict[str, Any]], skip: Optional[int], limit: Optional[int]) -> List[Dict[str, Any]]: Applies pagination to a list of serialized trackers using skip (offset) and limit parameters.

​

Configuration

tracker_store:
  type: path.to.your.module.Class
  url: localhost
  a_parameter: a value
  another_parameter: another value

​

Fallback Tracker Store