Changelog

We just shipped massive speed improvements across our platform! While we are continuously working on model speed, you'll notice everything else is running much faster now. Plus, we released some much-requested improvements on our chat input and API.

We've just added three powerful new models to Langdock: Claude 3.7 Sonnet, OpenAI's o3 Mini and Gemini 2.0 Flash.

Claude 3.7 Sonnet

Claude 3.7 Sonnet is the successor to 3.5, one of the most used models in our user base. The previous 3.5 version is used already by many users for writing tasks, such as emails or translations and for coding.

The upgrade of the new model version is a dual-mode capability, which were added as two separate models

• The normal mode allows users to either use it as a regular LLM and immediately create an answer for simpler tasks (like email generation or translating a text).
• The reasoning mode allows the model to self-reflect before answering, to provide a better, deeper answer for complex problems (like strategy, maths or science).

We have added the modes as two separate models (Claude 3.7 Sonnet and Claude 3.7 Reasoning).

o3 Mini

OpenAI's o3 Mini is the latest and most-efficient model of OpenAI's reasoning series.

Reasoning models, like o3 Mini, o1, r1 from DeepSeek or the Claude 3.7 Sonnet model mentioned above use chain-of-thought thinking to split a task into several steps. This makes them useful for complex tasks, like maths, physics, complex instructions, coding or complex strategic tasks.

o3 Mini is the latest and most efficient model of OpenAl's reasoning series. o1 is the broader knowledge reasoning model, and o3-mini is faster compared to the previous one, balancing speed and accuracy. As o3 Mini allows for control over its reasoning efforts, we have added the standard mode as well as a high-effort reasoning mode as two separate models (o3 Mini and o3 Mini High).

Gemini 2.0 Flash

We also added the new Gemini 2.0 Flash model, which is now available in the EU as well. The Flash model from the previous 1.5 Gemini generation was the faster, smaller model compared to the larger and more advanced Gemini 1.5 Pro. The new Gemini 2.0 Flash outperforms Gemini 1.5 Pro on key benchmarks and is twice as fast.

We are bringing a new way to interact with assistants in Langdock: Assistant forms. When building an assistant, editors can now choose to use the new form input method, where they can define the input fields shown to users.

You can build an interface to structure the inputs users need to enter to receive high-quality results, similar to survey forms. When users use an assistant using the new input method, they will be presented with the form the editor built. You can use inputs you know from other tools, like:

• Single-line text
• Checkboxes
• File upload
• Single-select options
• Number
• Date

This gives assistant creators more flexibility when creating assistants and allows them to tailor the input structure to your specific needs, while making it easier for other users to use the assistant.

Memory offers deeper personal customization of the model responses, by saving information from past interactions in the application.

When using memory, you can tell the model to remember certain information about you, your work or any preferences you have. It will then save the information in the application. For example, you could have it:

• Remember certain details about your job
• Share a preference for a specific style of writing
• Remember your name and other personal details

By default, Memory is disabled. To use it, head over to the preferences in your settings. There you can enable chat memory in the capabilities section.

All memories are stored in your account, and are available to you in all your chats (not assistant chats). They are not accessible by others in your workspace.

We've added support for the new R1 model from the Chinese AI company DeepSeek. R1 has been receiving a lot of attention in the media recently for its strong performance. The model rivals OpenAI's o1-series and is open-sourced for commercial use.

The R1 model is available in multiple versions. We are self-hosting the 32B version of the model on our own servers in the EU and consume the full 671B version from Microsoft Azure in the US. Since the model is still early and focuses on reasoning, we have deactivated tools like document upload, web search and data analysis for now.

Admins can enable the models in the settings.

We're excited to announce that you can now work with audio and video files in the chat.

Upload your recordings (up to 200MB) and our system will automatically transcribe them, allowing you to have natural conversations about the content.

You can work with all common formats including MP4, MP3, WAV, and MPEG files. Whether you need to review a team meeting, analyze a client call, or process a voice memo, simply upload your file and start asking questions about its content.

Langdock now offers an enhanced web search mode, providing quick and current answers along with links to relevant internal sources from the internet.

If web search is enabled for your workspace, you can now turn it on in the newly redesigned chat input bar. This will force the model to search the web for up-to-date news and information regarding your query.

We launched the API for assistants. You can now access assistants, including attached knowledge and connected tools through an API.

To enable an assistant to be accessible through the API, admins need to create an API key in the API settings. Afterward, you can share the assistant with the API by inviting the key like a normal workspace member.

After configuring the API in your workflow (here are our docs), you can then send messages to the assistant through the API. The API also includes structured output and document upload.

You can now share chats with users of your workspace by clicking on the button in the top right corner. It’s a great way to share your work with your colleagues directly in Langdock.

The sharing button appears after the first message has been sent in the chat. Once a chat is shared, new messages will also be shared. Others can only read the contents of that chat, but not interact with it. They will not have access to the documents attached, but can view answers based on the documents. You can unshare a chat anytime from the settings.

You can now navigate Langdock and search chats directly from your keyboard with the new command bar feature. This allows for quick and easy access to the information you need, right at your fingertips.

Pressing Cmd + K on your keyboard (for Windows Ctrl + K) opens a menu to quickly perform different operations. Here are a few examples:

• Search through all your chats
• Search a specific assistant
• Quickly change settings, like switching to dark mode, opening the documentation, the changelog, or the support chat

We also added a search button in the top left corner which opens the command bar.

There are also new and updated shortcuts:

• Open a new chat: Cmd/Ctrl + Shift + O
• Open/close the sidebar: Cmd/Ctrl + Shift + S
• Copy the last response: Cmd/Ctrl + Shift + C

We hope these improvements make you even more productive when using Langdock.

You can now incorporate variables directly into your prompts and create dynamic templates that can be easily reused across different contexts.

When creating a new prompt in the prompt library, wrap a word with {{ and }} or click on the variable button at the bottom to make it a variable. When using the prompt later, users can quickly fill out the variables in order to customize the prompt to their needs.

This helps to easily use a prompt in different contexts without leaving your keyboard or to make it easier for others to use the prompt when you share it.

You can find more details in our section about the prompt library in our documentation.

Gain valuable insights into how your assistants are being used with the new assistant usage insights feature available in Langdock.

Users can now upvote or downvote responses and leave comments, providing direct feedback that can help you improve your assistant's configuration. This interaction enhances the user experience and offers concrete suggestions for improvement in the feedback tab.

In the analytics tab, you as an editor or owner of an assistant, can access quantitative data about usage over specific timeframes in the analytics tab. The number of messages, conversations and users helps you understand user engagement and identify needs.

With these insights, assistant creators can assess performance and make informed improvements of the configuration, leading to a more effective and user-friendly assistant.

Introducing Canvas

We're excited to launch Canvas, a new feature in Langdock that enhances your writing and coding tasks. Canvas offers an interactive window where you can edit text and code directly, receive AI suggestions, and collaborate more effectively.

Highlights:

• Inline Editing: Easily make changes and get AI suggestions.
• Coding Tools: Review, and fix code seamlessly.
• Writing Enhancements: Adjust text tone and style with ease.

Canvas is now available for all Langdock users.

Assistants can now talk to other software tools like Jira or Salesforce via actions.

Actions

Assistants can now perform API calls to external tools, opening up many integration possibilities with CRMs, ticket management tools, or internal APIs. Check out our Actions documentation for details, including specialized guides for Jira, Google Drive, and Salesforce.

We added the latest OpenAI models, o1 and o1 mini. Admins can enable them in the model settings. As a heads up, these models are thinking models but not replacements for all tasks. The o1 models are better at reasoning and complex thinking tasks, like math, data analysis, or coding, than previous models, but not at knowledge retrieval, text generation, or translation. This is because they will take a comparatively long time to start writing their answer since they are thinking in the background first. You can read more about this in our model guide.