Changelog

New Features and Updates

• Unified Knowledge Bases

  • Manage all your knowledge bases in one centralized tab. Re-use them across projects without re-indexing files.
  • Now integrates with services like SharePoint or OneDrive without re-entering credentials.
  • Old connectors will still appear in the sidebar, but we suggest you start using this for better file navigation, RBAC (Role-Based Access Control), and re-usability across projects.

• Overall Analytics

  • Manage project usage in a single dashboard tab.
  • Options to sort by runs, users, tokens, and errors.

• Enhanced Charting

  • 10x improved charting capabilities for better aesthetics, more interactivity, and data filtering.
  • Download charts or open them in new tabs.
  • Check out this video for more details.

• Faster Web Search & Scraping

  • The new 'Web Search' node in Data Loaders is 10x faster, allowing you to search and scrape content efficiently.
  • Build your own Perplexity in minutes.

• Autosave

  • No more manual saves! Projects now automatically save with every change.
  • The last save time is displayed at the bottom of the screen.

• Improved Publish Experience

  • Name your versions when publishing for easy tracking and rollback.

• Run Progress Visualization

  • Visualize your workflow's progress in real-time.
  • Monitor inputs, outputs, latency, and more for each node, ensuring a clear understanding of every step.

New Meta Llama 3.2 models

Meta released Llama 3.2 models, now available in Stack AI, introducing several key advancements and new capabilities:

Multimodal Vision Models

Llama 3.2 introduces multimodal vision capabilities with two new models:

• 11B parameter model
• 90B parameter model These models can reason about high-resolution images, enabling tasks like image captioning, visual search, and interactive educational tools.

Lightweight Text Models

Two new lightweight text-only models have been introduced:

• 1B parameter model
• 3B parameter model These models are designed for edge and mobile devices, allowing for on-device AI experiences with minimal latency and resource usage.

Key Features

• Expanded Language Support: Improved multilingual capabilities covering 8 languages including English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
• Extended Context Length: All Llama 3.2 models support a 128K token context length.
• Improved Efficiency: The models offer reduced latency and improved performance for AI workloads.

Availability and Deployment

The Llama 3.2 models are now accessible in Stack AI via:

• Together AI
• Amazon Bedrock
• Meta dedicated node

New Features

• Loop Mode for External Flow Node
  We’ve introduced Loop Mode under the Utils section in the builder, allowing you to modularize your projects into individual components and execute them repeatedly in a loop. Perfect for tasks with variable inputs, such as competitive analysis, where an LLM generates a list of competitors that are then analyzed individually.

  • How to use Loop Mode: Simply format your input list as: ["input 1", "input 2", "input 3", ...]. Here’s a real-world example: A customer needed to identify competitors for a company. Using Loop Mode, they efficiently handled competitor analysis, ensuring the process could scale based on the number of competitors.

• Download Tables in Excel Format
  You can now download tables generated by your AI models in Excel format! Available in the following areas:

  • Builder: Through the Output Node
  • Form interface

• Connection Integration in Builder
  Now, you can use these connections directly in the builder with various knowledge bases and data loaders. Simply click the connection button to select or create a new connection on the spot.

New Models Added

• gpt-4o-2024-08-06:
  The latest snapshot supports Structured Outputs, ensuring responses follow your supplied JSON Schema—no more worries about missing keys or invalid enum values.
• chatgpt-4o-latest:
  This dynamic model continuously updates to the latest version of GPT-4o in ChatGPT, with frequent improvements.

New models available

• Llama 3.1 (via Groq, Together AI and Meta nodes):

  • World's largest and most capable open-source foundation model with 405 billion parameters (smaller versions are also available in Stack AI -> 70B and 8B for simpler tasks and faster execution)
  • Supports advanced use cases like long-form text summarization, multilingual conversational agents, and coding assistants, with a context length of 128K tokens.

• Mistral Large 2

  • 123 billion-parameter model with a 128K-token context window** (less parameters than Llama 3.1 an similar performance 🤯)
  • Excels in code generation, mathematics, and multilingual tasks, outperforming many leading models in these areas.

• GPT 4o mini

  • Cost-efficient small model that surpasses GPT-3.5 Turbo.
  • It offers superior textual intelligence and multimodal reasoning, with 128K-token context window.
  • Enables a broad range of tasks, including real-time text responses and applications requiring large context handling. Best for every day simple tasks

• Gemini 1.5 Flash

  • Lightweight and fast model optimized for speed and efficiency (comparable to GPT 4o mini)
  • Breakthrough one-million-token context window, making it ideal for processing extensive video, audio, and large codebases.

• Mistral nemo

  • Specialized model designed for high-precision tasks in scientific and technical domains, with a 128K-token context window.
  • It focuses on reducing hallucinations and improving reasoning capabilities, making it suitable for research and development applications.

Other features

• Images to LLMs with vision: upload or copy paste images in the chat assistant for processing by GPT 4o or Anthropic Claude 3.5 Sonnet.
• Google Search: select the country in which to perform the search.

New Feature

• Chart Generation in LLM Responses: Now you can enable chart generation directly through your LLM settings. Ideal for financial analysis, data exploration, and reporting, this feature allows for visually appealing charts rendered within the LLM's messages, enhancing data visualization and interpretation.

New Features

• Integration of Stable Diffusion 3: Medium, large, large turbo models.
• YouTube Node: Selection of transcription languages.
• Claude 3.5 Sonnet: Now available in AWS Bedrock.
• Qwen 2 72b: Integrated via Together AI.
• Charts in LLM Responses: Enable it in the settings of any LLM.

Other Enhancements

• URL Node: Revamped for a better scraping experience.
• Gmail Node: Now capable of reading email threads.

New Features

• Connections Manager

  • Create and manage connections in one place, eliminating the need to include credentials in every knowledge base for each project.
  • Add, refresh, remove, and test connections from the new Connections tab in Settings.
  • Choose from a list of available integrations in Stack AI and view details like creation date and status (e.g., healthy, unhealthy).

• Deepgram Nova 2 for Medical Transcriptions

  • 20.5% improvement in recognizing medical terms with Nova 2, which is 5-40x faster than SOTA models.
  • Other Available Submodels in the Nova 2 family: finance, meeting, voicemail, automotive
  • Healthcare Use Cases:
    • SOAP Notes: Automatically generate SOAP notes from patient meeting recordings.
    • Patient Intake: Efficient and accurate patient data capture.
    • Clinical Documentation: Improve the quality and consistency of medical records.
  • Read our blog post on AI solutions in Healthcare.

• Chat Assistant for OpenAI Downtime

  • Reliable Alternative: Use our own Chat Assistant at chat.stack-ai.com when OpenAI is down, powered by Anthropic, Groq, Gemini, Mistral, etc.
  • Consistent Functionality: Same functionality as ChatGPT, ensuring no performance issues.
  • Use this link

Other Enhancements

• Download Docx: Download responses from your AI form directly as a Docx.
• Batch Processing: Interface improvements for visualizing outputs and deleting all rows, available through the export tab.
• Chat Assistant: Fixes made to mobile version and delete functionality.
• Editable Submit Button: Customize the text on the submit button in the form export.

Other Important Announcements:

• SOC 2 Type II and HIPAA Compliance: We passed our SOC 2 Type II + HIPAA audit.
  • Contact us for the report at dpo@stack-ai.com.
  • Read more in our announcement post.

New Features

• GPT-4o on Azure

  • Dedicated Azure Instance: Access a dedicated Azure instance for GPT-4o, providing improved latency and security.
  • Easy Integration: Drag and drop an Azure node from the sidebar under the LLM section to boost your AI workflows.
  • API Key Integration: Integrate your own API keys if you have an Azure account.

• Text-to-SQL Enhancements

  • Interact with structured databases using natural language with our revamped Text-to-SQL model.
  • Supported Databases: PostgreSQL, Snowflake, MySQL
  • Improved model accuracy, response times, and error handling. Now supports larger schemas.
  • Read our latest blog post on building a Data Warehouse AI Assistant with the Text-To-SQL model and Snowflake.

• Voice Assistant in Beta

  • Build a voice assistant similar to the OpenAI GPT-4o demo using your proprietary data.
  • Integration Steps:
    • Drag and drop audio-to-text and text-to-audio nodes from the sidebar.
    • In the builder, go to the interface tab and select the audio assistant as your export option.
  • This feature is in beta, and we welcome your feedback.

Enhancements

• Video and Audio in RAG: Upload .mp4 and .mp3 files into your knowledge base for automatic transcription and integration.
• Multi-Language Support for Deepgram: Utilize audio-to-text features in various languages.
• Audio Upload in Forms: Upload audio files directly within the form interface.
• Download Button: Directly download responses from your AI applications via the form interface.
• Mistral Enhancements: All Mistral models now support JSON mode.
• Improved Web Scraping: Enhanced to remove subpages limitations for broader website scraping.
• Table + Search: New checks to automatically correct erroneous SQL queries.
• Summarize Node: Now powered by the OpenAI GPT-4o model for faster processing.
• Batch Processing: Interface improvements for easier batch processing.
• Chat Assistant: Enhanced with links in descriptions for easy access to relevant information.

Integrations

• OpenAI's Latest Model: GPT-4o ("o" for "omni")

  • Faster operations and 50% more cost-efficient in API usage.
  • Maintains high performance in English and coding tasks and enhanced in non-English text processing.
  • Multimodal capabilities to handle and output text, audio, images, and video.

• Groq

  • Groq is now accessible on Stack AI.
  • Specializes in Language Processing Units (LPUs) with ultra-low latency for AI inference.
  • Supports running open-source models like LLama-3 and Mixtral 8x7b efficiently.

Enhancements

• Interface Redesign
  • The 'Export tab' has been renamed 'Interface'.
  • All favorite interfaces (Chat, Forms, Embedded chatbots, WhatsApp/SMS, Slack, Batch, API) are now more accessible.
  • User-friendly configurations and real-time previews of modifications.