Rasa Pro Change Log
All notable changes to Rasa Pro will be documented in this page. This product adheres to Semantic Versioning starting with version 3.3 (initial version).
Rasa Pro consists of two deployable artifacts: Rasa Pro and Rasa Pro Services. You can read the change log for both artifacts below.
[3.13.10] - 2025-09-12
Rasa Pro 3.13.10 (2025-09-12)
Bugfixes
- Fixes the silence handling bug in Audiocodes Stream Channel where consecutive bot responses could trip silence timeout pattern while the bot is speaking. Audiocodes Stream channel now forcefully cancels the silence timeout watcher whenever it sends the bot response audio.
- Update
langchain
to0.3.27
andlangchain-community
to0.3.29
to fix CVE-2025-6984 vulnerability inlangchain-community
version0.2.19
.
[3.13.9] - 2025-09-03
Rasa Pro 3.13.9 (2025-09-03)
Bugfixes
- Upgrade
skops
to version0.13.0
andrequests
to version2.32.5
to fix vulnerabilities. - Fixed flow retrieval to skip vector store population when there are no flows to embed.
Miscellaneous internal changes
Miscellaneous internal changes.
[3.13.8] - 2025-08-26
Rasa Pro 3.13.8 (2025-08-26)
Improvements
- Validation errors now raise meaningful exceptions instead of calling sys.exit(1), producing clearer and more actionable log messages while allowing custom error handling.
- Enable connection to AWS services for LLMs (e.g. Bedrock, Sagemaker) via multiple additional methods to environment variables, for example by using IAM roles or AWS credentials file.
Bugfixes
- The .devcontainer docker-compose file now uses the new
docker.io/bitnamilegacy
repository for images, rather than the olddocker.io/bitnami
. This change was made in response to Bitnami's announcement that they will be putting all of their public Helm Chart container images behind a paywall starting August 28th, 2025. Existing public images are moved to the new legacy repository -bitnamilegacy
- which is only intended for short-term migration purposes. - Fix expansion of referenced environment variables in
endpoints.yml
during Bedrock model config validation. - Fixed model loading failures on Windows systems running recent Python versions (3.9.23+, 3.10.18+, 3.11.13+) due to tarfile security fix incompatibility with Windows long path prefix.
[3.13.7] - 2025-08-15
Rasa Pro 3.13.7 (2025-08-15)
Bugfixes
- Don't tigger a slot correction for slots that are currently set to
None
asNone
counts as empty value. - Enabled the data argument to support both string and list inputs, normalizing to a list for consistent handling.
[3.13.6] - 2025-08-08
Rasa Pro 3.13.6 (2025-08-08)
Bugfixes
- Fixed the contextual response rephraser to use and update the correct translated response text for the current language, instead of falling back to the default response.
- Refactored
validate_argument_paths
to accept list values and aggregate all missing paths before exiting.
[3.13.5] - 2025-07-31
Rasa Pro 3.13.5 (2025-07-31)
Bugfixes
-
Fix correction of slots:
- Slots can only be corrected in case they belong to any flow on the stack and the slot to be corrected is part of a collect step in any of those flows.
- A correction of a slot should be applied if the flow that is about to start is using this slot.
-
Upgrade
axios
to fix security vulnerability. -
Fix issue where called flows could not set slots of their parent flow.
[3.13.4] - 2025-07-23
Rasa Pro 3.13.4 (2025-07-23)
Improvements
- Engine-related modules now raise structured exceptions instead of calling
sys.exit(1)
orprint_error_and_exit
, providing clearer, more actionable log messages and enabling custom error handling.
Bugfixes
- Fix validation of the FAISS documents folder to ensure it correctly discovers files in a recursive directory structure.
- Updated
Inspector
dependent packages (vite
, and@adobe/css-tools
) to address security vulnerabilities. - Fixed bug preventing
model_group
from being used withembeddings
in generative response LLM judge configuration.
[3.13.3] - 2025-07-17
Rasa Pro 3.13.3 (2025-07-17)
Bugfixes
- Allowed NLG servers to return None for the text property and ensured custom response data is properly retained.
[3.13.2] - 2025-07-16
Rasa Pro 3.13.2 (2025-07-16)
Bugfixes
- Pass all flows to
find_updated_flows
to avoid creating aHandleCodeChangeCommand
in situations where flows were not updated. - Bugfix for Jambonz Stream channel Websocket URL path which resulted in failed Websocket Connections
[3.13.1] - 2025-07-14
Rasa Pro 3.13.1 (2025-07-14)
Bugfixes
- Fix issues with bot not giving feedback for slot corrections.
- Fixed bot speaking state management in Audiocodes Stream channel. This made the assistant unable to handle user silences
Added periodic keepAlive messages to deepgram, this interval can be configured with
keep_alive_interval
parameter
[3.13.0] - 2025-07-07
Rasa Pro 3.13.0 (2025-07-07)
Deprecations and Removals
- Deprecate IntentlessPolicy and schedule for removal in Rasa
4.0.0
. - Removed
monitor_silence
parameter from Voice Channel configuration. Silence Monitoring is now enabled by default. It can be configured by changing the value of Global Silence Timeout - Remove pre-CALM PII management capability originally released in Rasa Plus 3.6.0.
Remove
presidio
,faker
andpycountry
dependencies from Rasa Pro.
Features
-
Introducing the
SearchReadyLLMCommandGenerator
component, an enhancement over theCompactLLMCommandGenerator
. This new component improves the triggering accuracy ofKnowledgeAnswerCommand
and should be used when theEnterpriseSearchPolicy
is added to the pipeline. By default, this new component does not trigger theChitChatAnswerCommand
andHumanHandoffCommand
. Handling small talk and chit-chat conversations is now delegated to theEnterpriseSearchPolicy
.To incorporate the
SearchReadyLLMCommandGenerator
into your pipeline, simply add the following:pipeline:
...
- name: SearchReadyLLMCommandGenerator
... -
Implement Privacy Filter capability to detect PII in 3 supported events (SlotSet, BotUttered, and UserUttered) and anonymise PII from the event data.
The PII detection takes a tiered approach:
- the first tier represents a slot-based approach: the sensitive data is stored in a slot whose name is defined in the privacy YAML configuration.
- the second optional tier uses a default GliNer model to detect PII that is not captured by the slot-based approach.
The anonymisation of PII is done by redacting or replacing the sensitive data with a placeholder. These anonymisation rules are also defined in the privacy YAML configuration.
-
EnterpriseSearchPolicy
can now assess the relevancy of the generated answer. By default, the policy does not check the relevancy of the generated answer. But you enable the relevancy check by settingcheck_relevancy
totrue
in the policy configuration.policies:
- name: FlowPolicy
- name: EnterpriseSearchPolicy
...
check_relevancy: true # by default this is falseIf the relevancy check is enabled, the policy will evaluate the generated answer and determine if it is relevant to the user's query. In case the answer is relevant, the generated answer will be returned to the user. If the answer is not relevant, the policy will fallback to
pattern_cannot_handle
to handle the situation.pattern_cannot_handle:
description: |
Conversation repair flow for addressing failed command generation scenarios
name: pattern cannot handle
steps:
- noop: true
next:
... # other cases
# Fallback to handle cases where the generated answer is not relevant
- if: "'{{context.reason}}' = 'cannot_handle_no_relevant_answer'"
then:
- action: utter_no_relevant_answer_found
next: "END" -
Implement privacy manager class which manages privacy-related tasks in the background.
This class handles the anonymization and deletion of sensitive information in dialogue state trackers stored in the tracker store, as well as the streaming of anonymized events to event brokers. It uses background schedulers to periodically run these tasks and processes trackers from a queue to ensure that sensitive information is handled in a timely manner.
-
Add support for
jambonz_stream
voice-stream channel. Log level ofwebsockets
library is set toERROR
by default, it can be changed by the environment variableLOG_LEVEL_LIBRARIES
. -
Remove the beta feature flag check from the
RepeatBotMessagesCommand
. TheRASA_PRO_BETA_REPEAT_COMMAND
environment variable is no longer needed as the feature is GA in 3.13.0.
Improvements
-
Coverage report feature can now be used independently of RASA_PRO_BETA_FINE_TUNING_RECIPE feature flag.
-
Update default Embedding model from
text-embedding-ada-002
totext-embedding-3-large
. UpdateLLMBasedRouter
,IntentlessPolicy
andContextualResponseRephraser
default models to usegpt-4o-2024-11-20
instead ofgpt-3.5-turbo
. Update theEnterpriseSearchPolicy
to usegpt-4.1-mini-2025-04-14
instead ofgpt-3.5-turbo
. -
Update
MultistepCommandGenerator
to usegpt-4o
, specificallygpt-4o-2024-11-20
, instead ofgpt-3.5-turbo
asgpt-3.5-turbo
will be deprecated on July 16, 2025. Add deprecation warning forSingleStepCommandGenerator
; leave current default model asgpt-4-0613
. Adapt tests for CommandGenerators to usegpt-4o
instead ofgpt-3.5-turbo
due to upcoming model deprecation. UpdateLLMJudgeModel
andFine tuning Conversation Rephraser
to usegpt-4.1-mini
, specificallygpt-4.1-mini-2025-04-14
, instead ofgpt-4o-mini
, asgpt-4o-mini
will be deprecated on September 15, 2025. -
Make
CALM
template the default forrasa init
. -
Redis lock store now accepts
host
property from endpoints config. Propertyurl
is marked as deprecated for Redis lock store. -
Added support for basic authentication in Twilio channels (Voice Ready and Voice Streaming). This allows users to authenticate their Twilio channels using basic authentication credentials, enhancing security and access control for voice communication. To use this feature, set
username
andpassword
in the Twilio channel configuration.credentials.yamltwilio_voice:
username: your_username
password: your_password
...
twilio_media_streams:
username: your_username
password: your_password
...At Twilio, configure the webhook URL to include the basic authentication credentials:
# twilio voice webhook
https://<username>:<password>@yourdomain.com/webhooks/twilio_voice/webhook
# twilio media streams webhook
https://<username>:<password>@yourdomain.com/webhooks/twilio_media_streams/webhook -
Added two endpoints to the model service for retrieving the assistant’s default configuration and the default project template used by the assistant.
-
All conversations on Twilio Media Streams, Audiocodes Stream, Genesys channel ends with the message
/session_end
. This applies to both the conversation ended by user and the assistant. -
Refactor Voice Inspector to use AudioWorklet Web API
-
Added new CLI commands to support project-level linking and granular push/pull workflows for Rasa Studio assistants:
- Introduced
rasa studio link
to associate a local project with a specific Studio assistant. - Added
rasa studio pull
andrasa studio push
, with subcommands for granular resource updates (config
,endpoints
) between local and Studio assistants.
- Introduced
-
Implement
delete
method forInMemoryTrackerStore
,AuthRetryTrackerStore
,AwaitableTrackerStore
andFailSafeTrackerStore
subclasses. -
Implement
delete
method forMongoTrackerStore
. -
Implement
delete
method for SQLTrackerStore: this method accepts sender_id and deletes the corresponding tracker from the store if it exists. -
Implement
delete
method for DynamoTrackerStore: this method accepts sender_id and deletes the corresponding tracker from the store if it exists. -
Implement
delete
method forRedisTrackerStore
. -
Update
KafkaEventBroker
YAML config to accept 2 new parameters:stream_pii
: boolean (default: true). If set tofalse
, the broker will not publish un-anonymised events to the configured topic.anonymization_topics
: list of strings (default: []). If set, the broker will publish anonymised events to the configured topics when the PII management capability is enabled.
-
Add 2 new PII configuration parameters to
PikaEventBroker
:stream_pii
: Boolean flag to control whether or not to publish un-anonymised events to the configuredqueues
. If set to False, un-anonymised events won't be published to RabbitMQ. Defaults toTrue
for backwards compatibility.anonymization_queues
: List of queue names that should receive anonymized events with PII removed. Defaults to an empty list ([]
).
-
Add a new
anonymized_at
timestamp field to the Rasa Pro events supported by the PII management capability:user
slot
bot
This field indicates when the PII data in the event was anonymized. The field is set tonull
if the PII data has not been anonymized.
-
Add support for reading
privacy
configuration key from endpoints file to enable PII management capability. -
Add new
DELETE /conversations/<conversation_id>/tracker
endpoint that allows deletion of tracker data for a specific conversation. -
Cleaned up potential sources of PII from
info
,exception
anderror
logs. -
Allow default patterns to link to
pattern_chitchat
. -
Cleaned up potential sources of PII from
warning
logs. -
Add a warning log to alert users when exporting tracker data that contains unanonymized events.
-
Remove beta feature flag for pypred predicate usage in conditional response variations. Mark this functionality for general availability (GA).
-
Updated the default behavior of
pattern_chitchat
. With the deprecation ofIntentlessPolicy
,pattern_chitchat
now defaults to responding withutter_cannot_handle
instead of triggeringaction_trigger_chitchat
. -
PII deletion job now performs a single database transaction in the case of trackers with multiple sessions, where only some sessions are eligible for deletion. The deletion job overwrites the serialized tracker record with the retained events only. This ensures that the tracker store is updated atomically, preventing any potential tracker data loss in case of Rasa Pro server crashes or other issues during the deletion job execution.
Bugfixes
-
Fixes a bug where fine-tuning data generation raises a FineTuningDataPreparationException for test cases without assertions that end with one/multiple user utterance/s, as opposed to one/multiple bot utterance/s.
For these types of test cases, the fine-tuning data generation transcript now contains all test case utterances up to but excluding the last user utterance(s) as the test case does not specify the corresponding, expected bot response(s).
-
allowed for usage of litellm model prefixes that do not end in '<provider>/'
-
Fixes a bug in Inspector that raised a TypeError when serialising
numpy.float64
from Tracker -
- Fixes an issue with prompt rendering where minified JSON structures were displayed without properly escaping newlines, tabs, and quotes.
- Introduced a new Jinja
filter to_json_encoded_string
that escapes newlines (\n
), tabs (\t
), and quotes (\"
) for safe JSON rendering.to_json_encoded_string
filter preserves other special characters (e.g., umlauts) without encoding them. - Updated the default prompts for
gpt-4o
andclaude-sonnet-3.5
-
Fix intermittent crashes on Inspector app when Enterprise Search or Chitchat Policies were triggered
-
Add channel name to UserMessage created by the Audiocodes channel.
-
Reduced redundant log entries when reading prompt templates by contextualizing them. Added source component and method metadata to logs, and changed prompt-loading logs triggered from
fingerprint_addon
to useDEBUG
level. -
Prompts for rephrased messages and conversations are now rendered using agent, eliminating errors that occurred when string replacements broke after prompt updates.
-
Fix retrieval of latest tracker session from DynamoTrackerStore.
-
Files generated by the finetuning data generation pipeline are now encoded in UTF-8, allowing characters such as German umlauts (ä, ö, ü) to render correctly.
-
Fix an issue where the SetSlot and Clarify command value was parsed incorrectly if a newline character immediately followed the value argument in the LLM output.
-
Fixed the repeat action to include all messages of the last turn in collect steps.
-
- Fix an issue where running
rasa inspect
would always set theroute_session_to_calm
slot toTrue
, even when no user-triggered commands were present. This caused incorrect routing to CALM, bypassing the logic of the router. - Fix a regression in non-sticky routing where sessions intended for the NLU were
incorrectly routed to CALM when the router predicted
NoopCommand()
.
- Fix an issue where running
-
Fix validation and improve error messages for the documents source directory when FAISS vector store is configured for the Enterprise Search Policy.
-
Prevent slot correction when the slot is already set to the desired value.
-
Fallback to
CannotHandle
command when the slot predicted by the LLM is not defined in the domain. -
Fix tracker store PII background jobs not updating the
InMemoryTrackerStore
reference stored by the Agent. -
Fix potential
KeyError
inEnterpriseSearchPolicy
citation post-processing when the source does not have the correct citation. Improve the citation post-processing logic to handle additional edge cases. -
Fix issue with PrivacyManager unable to retrieve trackers from Redis tracker store during its deletion background job, because the Redis tracker store prefix was added twice during retrieval.
-
Fixed anonymization failing for
FloatSlots
when user-provided integer values don't match the stored float representation during text replacement. -
Fix Redis tracker store saving tracker with prefix added twice during the anonymization background job.
-
Fix DynamoDB tracker store overwriting tracker with the latest session in case of multiple sessions saved prior.
Miscellaneous internal changes
Miscellaneous internal changes.
[3.12.33] - 2025-09-12
Rasa Pro 3.12.33 (2025-09-12)
Bugfixes
- Fixes the silence handling bug in Audiocodes Stream Channel where consecutive bot responses could trip silence timeout pattern while the bot is speaking. Audiocodes Stream channel now forcefully cancels the silence timeout watcher whenever it sends the bot response audio.
- Fixed flow retrieval to skip vector store population when there are no flows to embed.
[3.12.32] - 2025-09-03
Rasa Pro 3.12.32 (2025-09-03)
Bugfixes
- Upgrade
skops
to version0.13.0
andrequests
to version2.32.5
to fix vulnerabilities.
Miscellaneous internal changes
Miscellaneous internal changes.
[3.12.31] - 2025-08-26
Rasa Pro 3.12.31 (2025-08-26)
Bugfixes
- The .devcontainer docker-compose file now uses the new
docker.io/bitnamilegacy
repository for images, rather than the olddocker.io/bitnami
. This change was made in response to Bitnami's announcement that they will be putting all of their public Helm Chart container images behind a paywall starting August 28th, 2025. Existing public images are moved to the new legacy repository -bitnamilegacy
- which is only intended for short-term migration purposes. - Fixed model loading failures on Windows systems running recent Python versions (3.9.23+, 3.10.18+, 3.11.13+) due to tarfile security fix incompatibility with Windows long path prefix.
[3.12.30] - 2025-08-19
Rasa Pro 3.12.30 (2025-08-19)
Bugfixes
- Don't tigger a slot correction for slots that are currently set to
None
asNone
counts as empty value.
[3.12.29] - 2025-07-31
Rasa Pro 3.12.29 (2025-07-31)
Bugfixes
-
Fix correction of slots:
- Slots can only be corrected in case they belong to any flow on the stack and the slot to be corrected is part of a collect step in any of those flows.
- A correction of a slot should be applied if the flow that is about to start is using this slot.
[3.12.28] - 2025-07-29
Rasa Pro 3.12.28 (2025-07-29)
Bugfixes
- Upgrade
axios
to fix security vulnerability. - Fix issue where called flows could not set slots of their parent flow.
[3.12.27] - 2025-07-23
Rasa Pro 3.12.27 (2025-07-23)
Bugfixes
- Fix validation of the FAISS documents folder to ensure it correctly discovers files in a recursive directory structure.
- Updated
Inspector
dependent packages (vite
, and@adobe/css-tools
) to address security vulnerabilities. - Fixed bug preventing
model_group
from being used withembeddings
in generative response LLM judge configuration.
[3.12.26] - 2025-07-17
Rasa Pro 3.12.26 (2025-07-17)
Bugfixes
- Allowed NLG servers to return None for the text property and ensured custom response data is properly retained.
[3.12.25] - 2025-07-16
Rasa Pro 3.12.25 (2025-07-16)
Bugfixes
- Pass all flows to
find_updated_flows
to avoid creating aHandleCodeChangeCommand
in situations where flows were not updated.
[3.12.24] - 2025-07-14
Rasa Pro 3.12.24 (2025-07-14)
Bugfixes
- Fixed the repeat action to include all messages of the last turn in collect steps.
- Fix issues with bot not giving feedback for slot corrections.
- Fixed bot speaking state management in Audiocodes Stream channel. This made the assistant unable to handle user silences
Added periodic keepAlive messages to deepgram, this interval can be configured with
keep_alive_interval
parameter
[3.12.23] - 2025-07-08
Rasa Pro 3.12.23 (2025-07-08)
Bugfixes
- Reverted fix for custom multilingual output payloads being overwritten.
[3.12.22] - 2025-07-07
Rasa Pro 3.12.22 (2025-07-07)
No significant changes.
[3.12.21] - 2025-07-03
Rasa Pro 3.12.21 (2025-07-03)
Bugfixes
-
Fixes a bug where fine-tuning data generation raises a FineTuningDataPreparationException for test cases without assertions that end with one/multiple user utterance/s, as opposed to one/multiple bot utterance/s.
For these types of test cases, the fine-tuning data generation transcript now contains all test case utterances up to but excluding the last user utterance(s) as the test case does not specify the corresponding, expected bot response(s).
-
Fix validation and improve error messages for the documents source directory when FAISS vector store is configured for the Enterprise Search Policy.
-
Prevent slot correction when the slot is already set to the desired value.
-
Fallback to
CannotHandle
command when the slot predicted by the LLM is not defined in the domain. -
Fix potential
KeyError
inEnterpriseSearchPolicy
citation post-processing when the source does not have the correct citation. Improve the citation post-processing logic to handle additional edge cases.
[3.12.20] - 2025-06-24
Rasa Pro 3.12.20 (2025-06-24)
Bugfixes
- Flows now traverse called and linked flows, including nested and branching called / linked flows. As a result, E2E coverage reports include any linked and called flows triggered by the flow being tested.
- Fixed a bug in Genesys and Audiocodes Stream channels where conversations used a placeholder value
default
as the Sender ID. Added a new methodVoiceInputChannel.get_sender_id(call_parameters)
that returns the platform'scall_id
as the Sender ID. This ensures each conversation has a unique identifier based on the call ID from the respective platform. Channels can override this method to customize Sender ID generation.
Miscellaneous internal changes
Miscellaneous internal changes.
[3.12.19] - 2025-06-18
Rasa Pro 3.12.19 (2025-06-18)
Bugfixes
- Fix issues where linked flows could not be cancelled and slots collected within linked flows could not be prefilled.
- Fix
InvalidFlowStepIdException
thrown when a bot is restarted with a retrained model which contains an update to the step order in a given active flow. Once retrained and restarted, the bot will now correctly handle the updated step order by triggeringpattern_code_change
.
[3.12.18] - 2025-06-12
Rasa Pro 3.12.18 (2025-06-12)
Bugfixes
- Ensure that old step ID formats (without the flow ID prefix) can be loaded without raising an
InvalidFlowStepIdException
in newer Rasa versions that expect the flow ID prefix. -
- Fix an issue where running
rasa inspect
would always set theroute_session_to_calm
slot toTrue
, even when no user-triggered commands were present. This caused incorrect routing to CALM, bypassing the logic of the router. - Fix a regression in non-sticky routing where sessions intended for the NLU were
incorrectly routed to CALM when the router predicted
NoopCommand()
.
- Fix an issue where running
- Enable slot prefilling in patterns.
[3.12.17] - 2025-06-05
Rasa Pro 3.12.17 (2025-06-05)
Bugfixes
- Fix an issue where the SetSlot and Clarify command value was parsed incorrectly if a newline character immediately followed the value argument in the LLM output.
[3.12.16] - 2025-06-03
Rasa Pro 3.12.16 (2025-06-03)
Bugfixes
- Make
domain
an optional argument inCommandProcessorComponent
. This change addresses a potentialTypeError
that could occur when loading a model trained without providing domain as a required argument. By making domain optional, models trained with older configurations or without a domain component will now load correctly without errors.
Miscellaneous internal changes
Miscellaneous internal changes.
[3.12.15] - 2025-06-02
Rasa Pro 3.12.15 (2025-06-02)
Improvements
:
Bugfixes
- Add turn_wrapper to count multiple utterances by bot/user as single turn rather than individual turns. Always include last user utterance in rephraser prompt's conversation history.
- Remove
StoryGraphProvider
from the prediction graph in caseIntentlessPolicy
is not present to reduce the loading time of the bot in cases where theIntentlessPolicy
is not used and a lot of stories are present.
[3.12.14] - 2025-05-28
Rasa Pro 3.12.14 (2025-05-28)
Improvements
-
- Updates the parameter name from
max_tokens
, which is deprecated by OpenAI, tomax_completion_tokens
. The oldmax_tokens
is not supported for theo
models. - Exposes LiteLLM's
drop_params
parameter for LLM configurations.
- Updates the parameter name from
Bugfixes
-
- Fixes default config initialization for the
IntentlessPolicy
.
- Fixes default config initialization for the
- Prompts for rephrased messages and conversations are now rendered using agent, eliminating errors that occurred when string replacements broke after prompt updates.
- Fix parsing of escape characters in translation of LLM prediction to SetSlotCommand.
- Files generated by the finetuning data generation pipeline are now encoded in UTF-8, allowing characters such as German umlauts (ä, ö, ü) to render correctly.
[3.12.13] - 2025-05-19
Rasa Pro 3.12.13 (2025-05-19)
Bugfixes
- The
Clarify
(syntax used bySingleStepLLMCommandGenerator
) /disambiguate flows
(syntax used byCompactLLMCommandGenerator
) command will now parse flow names with dashes. - Fix remote model download when models are stored in a path, not in the root of the remote storage.
Add new training CLI param
--remote-root-only
that can be used by the model service to store the model in the root of the remote storage. Propagate this parameter to the persistor'spersist
method. Simplify persistor code when retrieving models by downloading the model to the target path directly rather than copying the downloaded model to the target path. This also improved testability. - When inspector is not used, root server path should output:
Hello from Rasa: <version>.
. When inspector is used, root server path should output HTML page with a link to the path on which inspector can be reached. - Change the default value of
minimize_num_calls
in the config of LLM-based command generators toTrue
.
[3.12.12] - 2025-05-15
Rasa Pro 3.12.12 (2025-05-15)
Improvements
- Improved
CRFEntityExtractor
persistence and loading methods to improve model loading times.
Bugfixes
- Bumps aiohttp to 3.10.x, sentry-sdk to 2.8.x. Also bumps the locked versions for urllib3 and h11