跳转至

更新日志（归档）

0.0.32

  1. 新提供器：azure-openai - 添加了新的 Azure OpenAI Service 提供器。这允许您通过 Sample Connector for Bedrock 连接到 Azure OpenAI 部署。文档

0.0.31

  1. 新功能：跨账号负载和最大尝试次数设置 - 现在可以在 bedrock-converse 提供器中设置跨账号负载和最大尝试次数了。具体参见相关文档。

  2. BRConnector 模型更新 - 新部署的 BRConnector 中增加了如下模型：
    - Claude 3.5 Sonnet / v2 的跨区域模型
    - Claude 3.7 Sonnet 的模型

  3. CloudFormation 部署优化 - 优化了 CloudFormation 部署，现在结构更加清晰，提示更加友好了。

  4. Bugfix：修复嵌套自定义属性问题 - 修复了 sagemaker_lmi 提供器中的 x-amzn-sagemaker-custom-attributes 嵌套问题。

0.0.30

  1. 增强功能：添加 Think-Budget 动态支持 - 现在可以在 HTTP header 中指定 “Think-Budget”（值大于0）激活 API 动态调用思考模式（后端不要配置启用 thinking）。

  2. 新功能：sagemaker-lmi 提供器的自定义属性 - 添加了通过 header “X-Amzn-Sagemaker-Custom-Attributes” 向 sagemaker-lmi 提供器传递自定义属性的支持。例如：”X-Amzn-Sagemaker-Custom-Attributes: model0:/v1/DELETEions;model2:/xxx”。这个更新修复了之前把所有 headers 信息透传给 SageMaker 时，Custom-Attributes 最大长度超过 1024 的问题，注意：当前 X-Amzn-Sagemaker-Custom-Attributes 的最大长度仍然不能超过 1024 个字符。

  3. 修复：解决了非思考模式下的 thinkBudget bug - 修复了在非思考模式下与 thinkBudget 相关的问题。

0.0.29

  1. 增强功能：bedrock-converse 提供器现在支持 prompt cache 功能 - 现在可以在提供器后台配置 promptCache 了，请参考文档

  2. 增强功能：bedrock-knowledge-base 提供器现在可以直接输出 streaming 内容 - 之前 Bedrock 知识库没有 RetrieveAndGenere 的 streaming 版本，现在这个提供器支持由 Bedrock 知识库直接输出的 streaming 内容了。参考文档：https://aws-samples.github.io/sample-connector-for-bedrock/providers/bedrock-knowledge-base/

  3. 修复：费用更新的时候同时更新了最新时间 - 现在可以更新用户的 active 日期了，这样可以在 manager 后台看到最新活动的 api-key 了。

  4. Cloudformation 部署脚本更新 - 现在可以选择独立部署 vpc stack，然后再部署 BRConnector；ecs 部署增加了自动创建第一个 api key 的能力。

0.0.28

  1. 增强功能：流式模式下的 reasoning_content 和 tool_calls - 在流式（streaming）模式下，现在可以输出 reasoning_content 和 tool_calls 内容。解决了 issue #74 中提到的非流式模式下缺少 reasoning_content 字段的问题。

  2. 标准化：OpenAI 格式的函数调用 - 在流式和非流式模式下，函数调用（function calling）现在均适配了 OpenAI 的标准格式，提高了与现有工具和库的兼容性。文档

  3. 改进：费用统计的 total 现在直接修改成了 sql update，不会再出现月消费大于总消费了

  4. 新增：ECS 部署脚本 - 增加了 ECS 部署脚本。

  5. 修复：飞书消息重复 - 修复了飞书在 streaming 的情况下重复信息的问题。

0.0.27

  1. 新提供器：sagemaker-deepseek - 新增了部署在 SageMaker 上的 DeepSeek R1 模型提供商。该提供商支持带有 Deepseek 风格 reasoning_content 的流式输出。目前支持 SageMaker 上的三种部署方法：LMI、JumpStart 和 Bedrock。参考文档。

0.0.26

  1. 增强功能：bedrock-converse 工具调用 - bedrock-converse 现在输出 tool_calls 内容字段，支持使用 langchain.ChatOpenAI 请求进行工具调用。

  2. 改进：一致的消息 ID - 流式响应现在在单个请求的所有数据块中包含一致的消息 ID。

0.0.25

  1. Feature: Claude Sonnet 3.7 Thinking/Reasoning Support - Added support for thinking/reasoning capabilities with Claude Sonnet 3.7 on the bedrock-converse provider. For more details, refer to the documentation.

  2. Bugfix: Deepseek Format Reasoning Content - Fixed support for reasoning content in deepseek format within the built-in BRClient.

  3. Standardization: sagemaker-lmi finish_reason Value - Modified the finish_reason value in the sagemaker-lmi provider to consistently return “stop”.

0.0.24

  1. Bugfix: sagemaker-lmi finish_reason Output - This update addresses an issue where the finish_reason output from the sagemaker-lmi provider was returning an empty string. Now, it will correctly output a non-empty string, providing accurate information about the completion status.

  2. Feature: Display Model Reasoning in BRClient - Enhanced BRConnector to support and display the reasoning_content field from the new API, which shows the model’s step-by-step reasoning process. This update allows the frontend to properly recognize and render the model’s thought process in a think block format.

0.0.23

  1. New Provider: bedrock-deepseek – Added support for DeepSeek models deployed on Amazon Bedrock. This provider supports Think tags which are rendered as markdown blockquotes in the output. Detail.

0.0.22

  1. API Update: titan_embeddings – Updated the embeddings API to follow OpenAI-compatible format. The API now accepts both single string and array inputs through /v1/embeddings endpoint. Detail.

  2. Function-Calling Format Update: bedrock-converse – Changed the function-calling arguments parameter type from JSON to string to align with OpenAI’s format. Updated all related providers to ensure compatibility.

  3. Auto-login Feature for BRClient: Added an automatic login mechanism. If a user named __shared-api-key__ exists in the backend, the /portal-for-brclient.html entry point will automatically configure both the URL and key for BRClient. Note: This feature is intended for internal company use only. If __shared-api-key__ does not exist, only the API URL will be set, and the key must be manually configured in BRClient.

0.0.21

  1. New Provider: openai-compatible – This provider is designed to integrate APIs of models compatible with the OpenAI chat schema (tested with: deepseek, Doudao/Volcano platform, Qwen/Alibaba). Detail.
  2. New Provider: smart-router – This provider analyzes the semantics of user input and selects an appropriate model to query. Detail.
  3. New Model Configuration Mechanism: You can now configure a custom model to adapt to all models provided by a specific model provider. Detail.
  4. Bugfix: Corrected the cache loading logic, which previously caused a program crash upon the first run of BRConnector.
  5. Headers Value Pass-through in sagemaker-lmi Provider: Now, values in the headers are URL-encoded and included in the CustomAttributes tag of the SageMaker SDK.
  6. BRClient Update to Version 1.2.5: Key updates include:
    - In AKSK mode, users can now customize the Bedrock endpoint.
    - Fixed a bug where pressing Enter in Chinese input methods would send the message immediately.

0.0.20

  1. Fixed the representation of boolean types in environment variables. Now only “false” and “0” are considered as false.
  2. Added support for nova-canvas in the painter plugin (text-to-image only).
  3. Collapsed the display of operation buttons in the API key list.
  4. Modified the data type of the index in the output result choices to integer.
  5. Added a new Provider: nova-canvas, which allows image manipulation through natural language (including the following tools: text-to-image, image-to-image, foreground replacement/inpaint, background replacement/outpaint, background removal, variations, and color guidance). https://aws-samples.github.io/sample-connector-for-bedrock/zh/providers/nova-canvas/

0.0.19

  1. Fixed a critical bug: log file handles were filling up the system, causing it to become unresponsive after a period of time.
  2. Rewrote the logging system, aiming to unify the format as much as possible.

0.0.18

  1. Added model and API key queries to the backend management web interface.
  2. Added configuration for the Amazon Nova models in NoDB mode. As the Nova models are relatively new and only available in certain regions, please note to use them in conjunction with the AWS_DEFAULT_REGION environment variable.
  3. The nova model has been added to the initialization script in the database, but it will only take effect in the newly created databases of BRConnector.
  4. Fixed a backend error that attempted to save database data even in NoDB mode.
  5. Changed the output type of ‘id’ from number to string during streaming output, now it can be correctly deserialized by aider.
  6. Added support for the embedding API, now it can adapt to the titan embeddings model. You need to configure a new model using the titan_embeddings Provider, and use the API (/v1/embeddings) to call it.

0.0.17

  1. When the client actively stops output, the application now actively closes the server-side response.
  2. Cost control now works correctly.

0.0.16

  1. Added a new configuration option for bedrock-converse: maxTokens

0.0.15

  1. Added haiku 3.5 model in NoDB mode.

  2. Removed some unnecessary log outputs.

0.0.14

  1. Added caching mechanism, with the following details:

  • If the system does not configure environment variables PGSQL_HOST or PGSQL_DATABASE, BRConnector will switch to pass-through mode. In this mode, only the ADMIN_API_KEY environment variable can be used to access models, and only built-in static models can be accessed (default LLM models supported by Bedrock Converse; if new models are released in the future, BRConnector code needs to be updated).

  • If a database is configured and the PERFORMANCE_MODE environment variable is set, custom models and API Keys configured in the database will take effect. These two parts of data will be loaded directly into memory (up to 2000 entries, refreshed every 1 minute, significantly reducing database access). In this case, BRConnector will only verify if the API Key is valid, without checking if the API Key has permission to access a specific model. It will no longer save conversation records or track and control costs.

  1. Fixed a Dify access bug. Issue #43

  2. For Feishu webhook, added multi-turn dialogue, image recognition, and document recognition features.

  3. Added optional configurations for NoDB, PERFORMANCE_MODE in CloudFormation.

  4. Upgraded the aws-lambda-adapter image for Lambda to version 0.8.4. The new adapter version resolves conflicts with the Authorization Header in IAM_Auth for Lambda function URLs.

  5. Optimized CloudFormation Output, now showing clear endpoints.

0.0.13

  1. New Feature: Introduced a webhook mechanism for easy integration with third-party applications.

  2. New Webhook: Feishu, allowing BRConnector integration into Feishu through configuration.

  3. New Plugin: continue-coder, enabling BRConnector configuration in Continue for code generation.

  4. New Models Support: The painter plugin now supports the latest bedrock SD models: stability.stable-image-core-v1:0, stability.sd3-large-v1:0, and stability.stable-image-ultra-v1:0.

  5. Deployment Script Upgrade: Added a permission configuration, now correctly calling cross-region model profiles.

  6. Improvement: Added an API key query API, allowing precise matching by name, group_id, and role.

  7. Improvement: Default model support: Now a model named ‘default’ can respond to any model name. If not defined, an error message will be output.

  8. Refactored parts of the code and added a standard API: /v1/completions, to support continue-coder. Currently, only bedrock-coverse and continue-coder have implemented this API.

  9. Upgraded the underlying AWS SDK dependency to 3.645.0 and fixed various other bugs.

0.0.12

  1. New provider: urls-reader, which can parse URLs in user conversations and download the text content from the URLs to add to the conversation context.
  2. Optimized the painter provider, now supporting the Titan image model. The large language model has been changed to BRConnector’s local model.
  3. It is now possible to delete models, groups, and API Keys in the manager interface.
  4. Bug fix: Automatic update script for CloudFormation EC2.