Releases: aws-solutions/qnabot-on-aws
v6.1.0
[6.1.0] - 2024-08-29
Added
- Integration with Guardrails for Amazon Bedrock and Amazon Bedrock Knowledge Base Integration (see documentation)
- Ability to customize prompt template for RAG using Amazon Bedrock Knowledge Base through setting
KNOWLEDGE_BASE_PROMPT_TEMPLATE
(see documentation). - Ability to customize inference parameters for LLM specified in
BedrockKnowledgeBaseModel
inference parameters forBedrockKnowledgeBaseModel
through settingKNOWLEDGE_BASE_MODEL_PARAMS
(see documentation) - Ability to customize search type (e.g.
SEMANTIC
orHYBRID
) for how data sources in the knowledge base are queried through settingKNOWLEDGE_BASE_SEARCH_TYPE
(see documentation) - Ability to customize maximum number of retrieved results for RAG using Amazon Bedrock Knowledge Base through setting
KNOWLEDGE_BASE_MAX_NUMBER_OF_RETRIEVED_RESULTS
(see documentation). - Ability to customize metadata and filters for RAG using Amazon Bedrock Knowledge through setting
KNOWLEDGE_BASE_METADATA_FILTERS
(see documentation) - Added an option to specify the retention period for log groups through cloudformation parameter
LogRetentionPeriod
- Anonymized operational metrics for some designer settings
Changed
- Improved fault tolerance of Testall, Export, Import functionalities and added ContentDesignerOutputBucket
- Added Amazon Titan Text Embeddings V2 as an additional option to the list of embedding models provided through cloudformation parameter EmbeddingsBedrockModelId
- Added Amazon Titan Text Premier as an additional option to the list LLM models provided through cloudformation parameters LLMBedrockModelId and BedrockKnowledgeBaseModel. Issue 746
- Changed Sagemaker LLM image to latest
- Changed
CustomQnABotSettings
parameter store to Advanced Tier to accommodate storing additional custom settings
Removed
- Removed Amazon Lex V1 resources
- Removed Canvas LMS integration
Fixed
- Fixed import settings in content designer for double byte characters
- Fixed an edge case where the Knowledge Base could return a context starting with
#
characters, causing font differences in the returned text due to Markdown formatting - Fixed session attribute
qnabot_gotanswer
not being set totrue
after receiving hits from Knowledge Base
Security
- Security patch for axios, moto, read-excel-file, handlebars, boto3, click, elliptic & postcss
v6.0.3
[6.0.3] - 2024-08-06
Security
- Patched fast-xml-parser vulnerability
v6.0.2
[6.0.2] - 2024-07-22
Added
- Added a migration documentation for Migrating QnABot configurations and data from existing deployment to new deployment
- Added a documentation for Bedrock Knowledge Base
Fixed
- Improve logout functionality which signs out the user and invalidates the access and refresh tokens that Amazon Cognito issued to a user. Issue #747
- Fixed bug that restricted import of questions with answers that consisted of only double-byte characters. Issue #731
- Fixed bug with chained questions causing errors in the fulfillment lambda.
Updated
- Removed aws-sdk (JavaScript V2) from dependency list.
- Updated parameter description for elicit response bot settings in the content designer settings. Issue #745
- Removed LLM models
meta.llama2-70b-chat-v1
andmeta.llama2-13b-chat-v1
from the list of models in the Cloudformation parameterLLMBedrockModelId
since these models will be unavailable on Amazon Bedrock starting from August 12, 2024. - Updated the setting
LLM_QA_NO_HITS_REGEX
in the Content Designer to include a default patternSorry, I don't know
in prompts specified through the settingLLM_QA_PROMPT_TEMPLATE
and other patterns returned by LLMs in their responses. - Constrainted the query made to Bedrock Knowledge Base to maximum of 1000 characters input query as per the input requirements.
v6.0.1
[6.0.1] - 2024-06-26
Fixed
- Fixed bug that was restricting stack names to be below 26 characters. Issue #741
- Fixed a looping issue when using slots and chaining #721 (PR #721) - contributed by (@amendlik)
- Github links with incorrect paths.
Updated
- Security patches for braces, urllib3, and ws.
- Improved latency of IAM policy propagation when switching the Bedrock embedding model.
v6.0.0
[6.0.0] - 2024-06-07
Added
- Integration with Amazon Bedrock
- Integration with Amazon Bedrock Knowledge Base
- Enabled Fine Grain Access Control in Amazon OpenSearch Service by default. A new
OpenSearchFineGrainAccessControl
CloudFormation parameter has been added - Content Designer UI improvements (grouped fields)
- Anonymized operational metrics
- Enabled Kendra based authentication utilizing QnABot idToken. A new
AltSearchKendraIndexAuth
CloudFormation parameter has been added
(PR #513) - contributed by (@JasonHammett)
Updated
- Migrated AWS JavaScript SDK from v2 to v3 for Amazon Lex Web UI Integration
- Upgraded Amazon OpenSearch Service domain from 1.3 to 2.11 unlocking features such as snapshot management via OpenSearch Dashboards (for more information see Amazon OpenSearch release history)
- Renamed Elasticsearch to Opensearch and Kibana to OpenSearch Dashboards
- Migrated from ElasticSearch client to OpenSearch client
- Changed SageMaker LLM model from
Falcon-40b
toLlama-2-13b-chat
and Sagemaker embedding model intfloat/e5-large-v2 now uses the Jumpstart version of the model - Removed the cloudformation parameter
Encryption
to prevent disabling of Amazon OpenSearch Service domain and Amazon S3 encryptions - Updated Amazon OpenSearch Service domain minimum TLS policy to 1.2
- Rewrote and migrate S3 object deletion custom resources from JavaScript to Python to address 3rd party security vulnerability
- Updated CloudFormation template parameter groupings
- Kendra index configuration has been moved from Content Designer settings to CloudFormation parameters to restrict IAM permissions. The CloudFormation parameter
DefaultKendraIndex
has been replaced with three separate parameters:KendraWebPageIndexId
,KendraFaqIndexId
, andAltSearchKendraIndexes
. - Updated question validation when importing or creating questions in Content Designer
- Kendra Webcrawler will now create data source in the native language if supported by Kendra. Kendra can now query in different languages. (issue #713)
- Standardized folder structure
- Made logging enhancements for Amazon OpenSearch Service and Amazon Data Firehose
- Added case conversion handlebars helpers(PRs #719) - contributed by (@amendlik)
- Consolidated common AWS SDK dependencies to reduce overall solution artifact size and lambdas deployment package size
- Patched security vulnerabilities and and integrated with CloudFormation Guard evaluation tool
- Added ability to TestAll in a selected locale, the locales to be tested should be defined in the template parameter
LexV2BotLocaleIds
- Improved security for LLM inputs and outputs
Fixed
- Forgot password functionality for Content Designer login
- Chat client will now prompt the user after credentials are expired
- Specialty bot routing fixes, Lex Intent state fixes, handlebar processing of SSML injects "\n" into response (PRs #726) - contributed by (@bobpskier)
- Secured Cognito roles (PR #670) - contributed by (@amendlik)
- Improved error messages for embeddings (issue #665)
- AWS environment variable for region is wrongly set (issue #714)
- Fixed testall lambda timing out during status "Lex" when testing against large set of Q&As - contributed by (@jeve7)
- Fix Austrian German neural voice name (PR #729) - contributed by (@malte-aws)
- Fix npm update command (PR #681) - contributed by (@zherink)
- Fix build exits with code 1, if a CF template is invalid (PR #733) - contributed by (@richhaase)
v5.5.2
v5.5.1
[5.5.1] - 2024-04-01
Caution
Do not use this version because of an issue with the testall functionality which may introduce a high number of versions stored in the testall S3 bucket when Content Designer has no Q&As.
Fixed
- Document chaining was not working when using Lambda functions for chaining. This has been resolved. (issue #687)
- ESWarmer lambda was generating a big amount of log data in CloudWatch. This is now fixed. (issue #692)
Updated
- QnaBot Client to now use code grant instead of implicit grant for Cognito Authorization
- Security patch for webpack-dev-middleware
- Template to ensure an embedding instance size of 1 is at least choosen since serverless is no longer available for the embedding model
v5.5.0
[5.5.0] - 2024-01-04
Caution
Do not use this version because of an issue with the testall functionality which may introduce a high number of versions stored in the testall S3 bucket when Content Designer has no Q&As.
Added
- Added Core-Language parameter to the QnABot deployment. This parameter allows the user to select during the deployment a core language which will be used by the OpenSearch language analyzers to look for question and answers. With this update, QnABot can now be deployed natively in 33 Languages with a more syntactical accuracy for matching questions and answers
- Bot routing enhancements including passing initial utterance to specialty bot and receive session attributes from specialty bot (issue #376) - contributed by (@bobpskier)
- Improved error handling. Added custom error handling question to QnaUtility and some errors are appended to chat client message when ENABLE_DEBUG_RESPONSES is set to 'true'
- Added 'PROTECTED_UTTERANCES' setting which allows the user to configure a comma-separated list of utterances that will be ignored by LLM query disambiguation and translation. This fixes a bug where feedback (thumbs up/thumbs down) and language selection would be disambiguated instead of triggering the respective workflow
- Added 'getQuestion' handlebar that returns the original matched question without hard-coding (issue #397)
- Added functional test collection for verifying deployed QnABots
- Added Service API Usage Tracking
- Added deployment parameter to enable selection of opensearch instance type (issue #599)
Updated
- Migrated out of Bluebird promises to native promises
- Migrated to AWS SDK for JavaScript v3
- Upgraded to Webpack 5
- Upgraded to Vue3
- Upgraded to Vuetify 3
- Upgraded to latest LLM Image
- Code Quality improvements based on SonarQube analysis
- Security patches for npm
Fixed
- Fixed chaining not working when combined with bot routing (issue #508) - contributed by (@bobpskier)
- Fixed issue with chaining causing QnABot to become unresponsive when chaining rule evaluation fails. Improved error reporting when debugging is enabled.
- Fixed issue preventing lambda hooks defined in the templates/extensions directory from being executed by the fulfillment lambda.
- Fixed issue where LLM errors return 'no_hits' response instead of error message.
- Fixed bug where positive feedback is not published to SNS.
- Fixed content designer settings using different casing standard for boolean values (issue #666)
- Fixed inclusion of OpenSearch QnA results in text passages (issue #669) - contributed by (@cristi-constantin)
- Fixed issue where session attributes become undefined when translate isn't enabled.
- Fixed issue where settings were being evaluated as strings instead of numbers. Settings that are saved as stings that represent positive, negative, whole, or decimal numbers will be parsed as numbers.
- Fixed issue where kendra redirect does not use redirect query when users locale matches kendra index locale
v5.4.5
[5.4.5] - 2023-11-1
Caution
Do not use this version because of an issue with the testall functionality which may introduce a high number of versions stored in the testall S3 bucket when Content Designer has no Q&As.
Updated
- Security patch for browserify-sign
v5.4.4
[5.4.4] - 2023-10-24
Caution
Do not use this version because of an issue with the testall functionality which may introduce a high number of versions stored in the testall S3 bucket when Content Designer has no Q&As.
Updated
- Security patch for urllib3
Fixed
- Fixed fulfillment throws an exception when a Preprocessing Lambda Hook is configured(issue #651)
- Improved error handling