Projects

Overview: DocPaws is an engineering-oriented RAG document assistant that manages knowledge bases, indexes PDFs, supports stream-based dialogues, and handles scope-based retrieval and citation. It is built with FastAPI and Vue3, and features a comprehensive architecture document and detailed backend and frontend documentation. Key features: - Personal knowledge base, folder uploads, and manifest-based incremental indexing - Stream-based RAG conversations (scope: full library, folder, single file), stream SSE, and thought process visualization - Retrieval distance threshold rejection (RETRIEVAL_MAX_DISTANCE) for questions beyond the threshold - Golden 20 reproducible evaluation Tech stack: Python, FastAPI, Vue3, SQLite, PostgreSQL, Redis, Docker, Node.js, Vite Why it matters: DocPaws simplifies the process of managing and retrieving information, making it easier for developers to find and reference documentation efficiently. Its robust architecture and comprehensive documentation ensure reliability and ease of use. Repository: biao994/DocPaws GitHub stars: 181 GitHub description: 工程化 RAG 文档助手：知识库、PDF 索引、流式对话、引用与检索阈值拒答。FastAPI + Vue3

Overview: This project focuses on enhancing the perceptual quality of video super-resolution (VSR) using Diffusion Models (DMs) while ensuring temporal consistency among frames. The method, named StableVSR, leverages a Temporal Conditioning Module (TCM) to synthesize realistic and temporally-consistent details. Key features: - StableVSR: A VSR method based on Diffusion Models that enhances perceptual quality. - Temporal Conditioning Module (TCM): Guides the generative process to produce high-quality and temporally-consistent results. - Frame-wise Bidirectional Sampling: Encourages the use of information from past to future and vice-versa to improve perceptual quality and temporal consistency. Tech stack: Python Why it matters: StableVSR addresses the need for improved perceptual quality and better temporal consistency in video super-resolution, making it a significant advancement in the field of image and video processing. Repository: claudiom4sir/StableVSR GitHub stars: 178

### Overview: Agentic Signal is a visual AI workflow automation platform that integrates local AI intelligence, enabling developers to build intelligent workflows without the need for cloud dependencies. ### Key Features: - **Visual Workflow Builder**: Use a drag-and-drop interface to create workflows. - **Node-Based Architecture**: Connect data sources, AI processors, and outputs seamlessly. - **Real-time Execution**: See workflows run in real-time with live data flow. - **Local AI Intelligence**: Run AI models locally with privacy and control. - **Tool Calling**: AI agents can execute functions and access external APIs. - **Structured Output**: JSON schema validation for reliable AI responses. - **Conversation Memory**: Maintain context across workflow executions. ### Tech Stack: TypeScript ### Why It Matters: Agentic Signal simplifies the process of building intelligent workflows, making it easier for developers to automate tasks without the need for cloud services. This platform leverages local AI intelligence, ensuring privacy and control over data and workflows.

### Overview The Common Voice Datasets repository contains metadata and release details for the Common Voice project. This project collects voice data through various modalities and provides detailed information about each dataset type, including scripts, spontaneous speech, and code switching. The datasets are available for download from the Mozilla Data Collective (MDC) platform. ### Key Features - **Scripted Speech (SCS)**: Contains scripted speech data, with a total of 43,000 hours of validated data. - **Spontaneous Speech (SPS)**: Includes spontaneous speech data, with a total of 600 hours of validated data. - **Code Switching (CS)**: Provides data on code switching, with a status of Alpha. ### Tech Stack The project uses JavaScript for its technical implementation. ### Why It Matters The Common Voice Datasets are crucial for researchers, linguists, and language enthusiasts who need access to high-quality voice data for language analysis and development. The datasets are widely used in various academic and research projects, contributing to advancements in language technology and linguistics.

Overview: This project is designed to guide Java developers through the transition into becoming AI engineers, focusing on the Spring AI framework and integrating powerful language models like OpenAI. Key features: - Spring AI Official Documentation - OpenAI Platform Docs - Ollama - AWS Bedrock - Docker Desktop - Docker Model Runner - Qdrant Vector Database - Model Context Protocol (MCP) - Prometheus - Micrometer - OpenTelemetry - Grafana - Jaeger Tracing Tech stack: Java Why it matters: This repository provides comprehensive resources and tools for developers transitioning into AI engineering, focusing on practical applications of Spring AI and integrating AI models with Spring Boot applications.

Overview: This repository contains various applications of Generative AI and Large Language Models (LLMs), including a voice assistant chatbot, an AI squad for automating tasks, and a chatbot based on multimodal data. Key features: - Build a voice assistant LLM chatbot using Ollama and speech recognition - Create an AI squad for automating tasks on your laptop without a GPU - Build a chatbot using multimodal data for better performance Tech stack: Jupyter Notebook Why it matters: This repository provides practical examples of how Generative AI and LLMs can be applied in various contexts, demonstrating their capabilities and potential use cases.

### Overview: ChatData is a chat-with-documents application that leverages the MyScale free knowledge base and LangChain's Self Querying Retriever to provide answers to natural language queries. It features three key features: a self-querying retriever, vector SQL for complex queries, and session management to maintain user interactions. ### Key features: - **Self-querying retriever**: Enhanced with the ability to use timestamps and arrays of strings for more data types. - **Vector SQL**: SQL optimized for vector databases, allowing for complex search queries. - **Session Management**: Enables seamless continuation of interactions with previous sessions. ### Tech stack: Python ### Why it matters: ChatData provides a powerful and efficient way to access vast amounts of knowledge, making it ideal for researchers, students, and anyone looking to explore academic papers and research documents. Repository: myscale/ChatData Primary language: Python GitHub stars: 178

Overview: 本项目是一份全面的AI工程师成长手册，涵盖从基础知识到RAG、Agent实战的系统学习路径，以及高频面试题的准备资料。 Key features: - 系统梳理AI基础知识 - 精选AI学习资料推荐 - AI常见名词速查表 - 从零构建检索增强生成系统 - 详细讲解RAG核心链路 - 深入解析向量检索与生成 - 解决工具调用与MCP协议问题 Tech stack: Python、Markdown、HTML、CSS、JavaScript Why it matters: 为程序员转型AI工程师提供系统学习路径和实战经验，帮助他们从零开始掌握AI技术，准备面试并实现项目落地。

Overview: This repository is a curated list of AI PowerPoint tools, PowerPoint automation, PPTX editing, and slide workflow projects. Key features: - HTML-style PowerPoint solutions - Image-generation PowerPoint solutions - PPTX library generation solutions - PPTX library and automation infrastructure Tech stack: Python Why it matters: It provides a comprehensive list of tools and projects that can help developers create, edit, convert, or check PowerPoint presentations. Repository: ningzimu/awesome-ai-ppt GitHub stars: 174

### Overview: RAGGENIE is an open-source, low-code platform designed to simplify the creation of custom Retrieval-Augmented Generation (RAG) chatbots with your own data. The project is currently in its early stages and aims to make AI development accessible and efficient. ### Key Features: - **RAG Builder**: Easily build conversational AI applications using RAG technology. - **Data Integration**: Connect to multiple data sources to enrich your chatbot with relevant information. - **Custom Capabilities**: Define and execute custom functionalities for your chatbot. ### Tech Stack: Python ### Why it Matters: RAGGENIE simplifies the process of creating custom chatbots, making it accessible to developers without extensive technical knowledge. The platform's open-source nature allows for easy community contributions and customization, ensuring continuous improvement and support.

Overview: PromptingTools.jl is a Julia package designed to simplify interactions with large language models, aiming to save developers 20 minutes per day. Key features: - Streamlines interactions with large language models - Easy templating for dynamic responses - Supports various models and APIs - Integrates with Prepaid Billing for cost management Tech stack: Julia Why it matters: This package is essential for developers looking to automate tasks and improve productivity by reducing the time spent on interacting with AI models. Repository: svilupp/PromptingTools.jl GitHub stars: 177 GitHub description: Streamline your life using PromptingTools.jl, the Julia package that simplifies interacting with large language models.

### Overview: UniversalRAG is a novel retrieval-augmented generation framework designed to handle corpora of diverse modalities and granularities. It introduces a modality-aware routing mechanism that dynamically identifies the most appropriate modality-specific corpus for each query, addressing the limitations of modality gaps and fixed-granularity retrieval. ### Key features: - **Modality-Aware Routing Mechanism**: Dynamically identifies the most suitable modality-specific corpus for each query. - **Retrieval-Augmented Generation**: Enhances the quality of generated text by leveraging retrieval results. - **Versatile Granularities**: Supports granular retrieval across different levels of granularity. - **Cross-Modal Integration**: Seamlessly integrates text, image, and video modalities for comprehensive processing. ### Tech stack: Python ### Why it matters: UniversalRAG addresses the limitations of traditional retrieval-based models by integrating the strengths of generation and retrieval, making it particularly effective for tasks requiring diverse and granular data handling. Its modality-aware routing mechanism ensures that the most relevant and appropriate data is used for generating high-quality output, making it a valuable tool for researchers and practitioners in the field of natural language processing and information retrieval.

### Overview: Hash3D is a universal solution designed to accelerate score distillation sampling (SDS) based 3D generation tasks, making the process significantly faster without the need for training. ### Key features: - **In-place Acceleration**: Hash3D improves the speed by up to 4 times compared to traditional methods. - **Training-Free**: No need for training; it directly accelerates the inference process. - **Compatibility**: Compatible with any 3D generation method using SDS. ### Tech stack: Python ### Why it matters: By effectively hashing and reusing feature maps across neighboring timesteps and camera angles, Hash3D substantially prevents redundant calculations, significantly accelerating the diffusion model's inference in 3D generation tasks.

### Overview: LoRA (Low Rank Adaptation) implemented in Rust for use with Candle. This technique interchanges the fully-trainable layers of the model with new, LoRA layers, allowing for more efficient fine-tuning. Weight merging improves inference performance by nullifying the added cost of using a fine-tuned LoRA model. ### Key features: - Converts `Linear`, `Conv1d`, `Conv2d`, `Embedding` layers into LoRA layers. - Implements weight merging to enhance inference performance. - Provides weight unmerging. - Offers easy-to-use APIs. - Uses an extensible trait-based layer swapping mechanism. ### Tech stack: Rust ### Why it matters: This implementation allows for more efficient fine-tuning while maintaining performance, making it suitable for various applications requiring model adaptation.

### Overview: Bello is a powerful tool designed to optimize the usage of your existing AI models, enhancing the quality of your coding tasks without increasing costs. It integrates a supervisor, a reviewer, an adversary, and a local log distiller around the coder, each with its own model, to provide a more efficient and reliable coding environment. ### Key features: - **Configuration X**: Offers the best balance of price, quality, and time for your coding task. - **Log Distiller**: Reduces API-equivalent cost by 22% on Astra and 14.8% on Luna. - **Adversary**: Ensures the coder's work is reviewed for accuracy and completeness. - **Completion Reviewer**: Helps in identifying and fixing any errors in the coder's output. - **Runtime Supervisor**: Provides guidance and support during the task execution. ### Tech stack: Python ### Why it matters: Bello is particularly useful for developers who want to leverage their existing AI models while maintaining high quality and efficiency. By integrating a variety of models and components, Bello ensures that the coder receives the best possible support and feedback, without compromising on the quality of the output. This makes Bello an indispensable tool for anyone looking to streamline their coding process and achieve better results with their AI resources.

Overview: Dive into the cutting-edge world of diffusion-based generative AI, learning to build your own model from scratch. Key features: - Explore the cutting-edge world of diffusion-based generative AI and create your own diffusion model from scratch. - Gain deep familiarity with the diffusion process and the models driving it, going beyond pre-built models and APIs. - Acquire practical coding skills by working through labs on sampling, training diffusion models, building neural networks for noise prediction, and adding context for personalized image generation. - At the end of the course, you will have a model that can serve as a starting point for your own exploration of diffusion models for your applications. Tech stack: Jupyter Notebook Why it matters: This course is essential for developers looking to expand their generative AI capabilities, providing a comprehensive understanding of diffusion models and practical coding skills needed to build and optimize your own models.

Overview: 本项目提供了一个从零开始学习大模型（LLM）的教程，通过手写GPT-2模型，详细讲解了其背后的原理和实现过程。 Key features: - 从零手写GPT-2模型，包含PyTorch代码 - 分词、嵌入、注意力机制、多头注意力、残差连接、层归一化等核心模块的详细实现 - 用"猫吃鱼"的例子，逐行拆解推理过程，每一步都配有代码实现和可运行脚本 - 和HuggingFace官方实现逐项对齐，确保结果一致 Tech stack: Python, PyTorch Why it matters: 本教程适合零基础的Python开发者，通过实际操作，帮助读者理解大模型的工作原理，无需依赖现成的Transformer模块，从头学到训练和推理的全过程。 Repository: aa1143/NamiGPT GitHub stars: 143

Overview: FastEmbed-js is a TypeScript/NodeJS implementation of the popular Qdrant FastEmbed library, designed for generating vector embeddings. It supports CommonJS and ESM, and utilizes native bindings for tokenizers. The default model is Flag Embedding, ranking among the top in the MTEB leaderboard. Key features: - Supports batch embeddings with generators. - Uses multi-arch native bindings for Huggingface/tokenizers. - Initializes models like BAAI/bge-base-en, BAAI/bge-base-en-v1.5, and others. - Provides passage and query embeddings for enhanced accuracy. Tech stack: TypeScript Why it matters: FastEmbed-js offers a fast, lightweight, and accurate solution for embedding text data, making it suitable for various applications requiring vector-based text representations. GitHub stars: 177 Repository: anush008/fastembed-js

### Overview This is a comprehensive framework designed to enable Generative BI capabilities on customized data sources (RDS/Redshift) hosted on AWS. It offers the following key features: - Text-to-SQL functionality for querying customized data sources using natural language. - User-friendly interface for adding, editing, and managing data sources, tables, and column descriptions. - Performance enhancement through the integration of historical question-answer ranking and entity recognition. - Customize business information, including entity information, formulas, SQL samples, and analysis ideas for complex business problems. - Add agent task splitting function to handle complex attribution analysis problems. - Intuitive question-answering UI that provides insights into the underlying Text-to-SQL mechanism. - Simple agent design interface for handling complex queries through a conversational approach. ### Key Features - Text-to-SQL functionality - User-friendly interface - Performance enhancement - Custom business information - Agent task splitting - Question-answering UI ### Tech Stack The tech stack includes: - Amazon Bedrock - Amazon OpenSearch Service ### Why It Matters This solution provides a practical approach to generating business insights from text data, making it accessible and user-friendly. The integration of Amazon Bedrock and Amazon OpenSearch with RAG technique ensures efficient and accurate analysis, making it valuable for businesses looking to automate their BI processes.

### Overview: RAFT, or Retrieval-Augmented Fine-Tuning, is a method that combines fine-tuning and retrieval-based querying to create realistic agents that can convincingly mimic a specific human target. This approach is particularly suited for applications requiring high-fidelity dialogue and conversation simulation. ### Key features: - **Fine-tuning**: Trained models on a specific dataset to achieve high accuracy. - **Retrieval-based querying**: Utilizes a knowledge graph to retrieve relevant information from a database, enhancing the agent's ability to answer questions and provide contextually relevant responses. - **CLI Options**: The repository includes command-line interfaces for fetching, chunking, embedding, summarizing, and querying a database, making it easy to integrate into various projects. ### Tech stack: Python ### Why it matters: This repository serves as a foundational tool for developing conversational agents and chatbots that can convincingly simulate human behavior, particularly in fields such as artificial intelligence and natural language processing. The integration of retrieval-based querying allows for more dynamic and contextually rich interactions, making it valuable for applications requiring high-fidelity dialogue and conversation simulation.