IN TODAY'S SIGNAL

🎖️ Top News

• Master unstructured data using RAGFlow for accurate, coherent content.

📌 Retool

• Read The State of AI Report: from 700+ tech leaders on AI ROI and key use cases.

⚡️ Trending Signals

• Runway opens Gen-3 Alpha: most powerful text-to-video model
  
  

• HuggingFace adds RT-DETR: real-time detection transformer
  
  

• Google opens Gemini 1.5 Pro: 2M token context for all devs
  
  

• Dev team enhances Jupyter with open-source code-generation models
  
  

• Beating NumPy's matrix multiplication in 150 lines of C code

🤗 Top of HuggingFace

• MARS5-TTS generates speech: 5 seconds of audio, text snippet, diverse scenarios
  
  

• Google's Gemma-2-9B: lightweight, state-of-the-art models, built on Gemini tech
  
  

• Florence-2-large: advanced vision model for diverse vision-language tasks
  
  

• Infinity-Instruct dataset: 10 million high-quality instructions to fine-tune and enhance models
  
  

• Cambrian-10M dataset: designed for multimodal instruction tuning, visual interaction data
  
  

• PersonaHub dataset: 1 billion diverse personas curated from web data

🧠 Tutorial

• The Ultimate Guide for Mastering Retrieval-Augmented Generation (RAG)

TOP NEWS

RAG

⇧ 10,711 Stars

What's New

Retrieval-Augmented Generation (RAG) is a method that combines the capabilities of information retrieval and language generation. This approach enhances the quality of generated content by first retrieving relevant information from a large dataset and then using that context to generate responses. The RAG model ensures that the answers or content produced are not only coherent but also factually accurate and based on existing data.

Capabilities of RAGFlow

RAGFlow is an open-source engine that implements the RAG methodology, focusing particularly on deep document understanding. Here’s what RAGFlow allows users to do:

Integrations

• Integration with LLMs such as OpenAI GPT-4o, DeepSeek-V2, Baichuan, and VolcanoArk.
  
  
• Enhanced text retrieval capabilities through the addition of BCE and BGE reranker models.
  
  
• New support for Markdown and Docx in the Q&A parsing method, along with capabilities for extracting images and tables from these formats.

Features

• Deep Document Understanding: Extracts knowledge from unstructured data in complex formats, ensuring high-quality input leads to high-quality output.
  
  
• Advanced Data Retrieval: Capable of locating specific information within a vast array of data, handling virtually unlimited tokens.
  
  
• Template-Based Chunking: Offers a range of templates for structured document parsing, combining intelligence with explainability.
  
  
• Grounded Citations: Minimizes errors in data retrieval by providing verifiable citations, with visualization tools that permit human intervention during text chunking.
  
  
• Broad Compatibility: Works with a diverse array of data sources including Word documents, PowerPoint slides, Excel sheets, text files, images, scanned documents, structured data, and web pages.
  
  
• Streamlined Workflow: Automates the RAG process, making it efficient for both individual and enterprise-scale applications.
• Configurable Models: Allows for customization of Large Language Models (LLMs) and embedding models to suit specific needs.

Access

• Use the RAGFlow Demo

• Use the Github repo to run it locally

Retool’s newest State of AI report just launched, and it’s packed with data from 700+ devs and tech leaders to help you cut through the hype and learn how to leverage AI for real impact.

The report covers:

- The good, the bad, and the ugly of the AI stack

- How much ROI builders are actually seeing from their AI use

- The real AI use cases builders and businesses love most

TRENDING SIGNALS

Video Generation

⇧ 2810 Likes

Detection Transformer

⇧ 839 Likes

Language Models

⇧ 1110 Likes

Notebooks

⇧ 46 Likes

Implementation

⇧ 163 Likes

TOP OF HUGGINGFACE

Models

• MARS5-TTS: With just 5 seconds of audio and a snippet of text, MARS5 can generate speech even for prosodically hard and diverse scenarios like sports commentary, anime and more.
  
  

• gemma-2-9b: Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models.
  
  

• Florence-2-large: Florence-2 is an advanced vision foundation model that uses a prompt-based approach to handle a wide range of vision and vision-language tasks.

Datasets

• Infinity-Instruct: Infinity Instruct contains 10 million high-quality instructions to fine-tune models and enhance performance.
  
  

• Cambrian-10M: Cambrian-10M is a comprehensive dataset designed for instruction tuning, particularly in multimodal settings involving visual interaction data.
  
  

• PersonaHub: a collection of 1 billion diverse personas automatically curated from web data. These 1 billion personas (~13% of the world's total population)

TUTORIAL

RAG

⇧ 293 Likes

LAST WEEK'S GREATEST HITS

• Anthropic launches Claude 3.5: surpasses GPT-4o, 2x faster, 80% cheaper than Opus.
  
  

• Meta introduces new open-source models for vision, text, and watermarking.
  
  

• OpenAI cofounder, Ilya Sutskever, starts Safe Super-Intelligence company.