Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 213 stars

shibing624/TreeSearch

TreeSearch: Search your codebase like a human — not like a vector database. No embeddings. No chunking. Just millisecond search over structured documents and large codebases. 无需 embedding，无需切分文档，在结构化文档和大型代码库中实现毫秒级检索。

Overview: TreeSearch is a structure-aware document retrieval library designed for efficient and precise search over structured documents and large codebases. It offers millisecond-latency performance without the need for vector embeddings or document chunking. Key features: - Supports structured documents and large codebases - Implements millisecond-latency search over tens of thousands of documents and large codebases - Preserves structure while performing searches - No need for vector embeddings or document chunking Tech stack: Python Why it matters: TreeSearch simplifies codebase search by leveraging structured documents and large codebases without the complexities of vector embeddings or document chunking, making it ideal for developers looking for a more intuitive and efficient search experience. Repository: shibing624/TreeSearch GitHub stars: 213 GitHub description: TreeSearch: Search your codebase like a human — not like a vector database. No embeddings. No chunking. Just millisecond search over structured documents and large codebases.

Java 213 stars

t1804330987/DD_Rag

基于 SpringAI 的 Agent 开发项目：一个面向“组织知识库 + AI 助手”的 RAG Agent实战项目，把权限隔离、文档入库、混合检索、证据约束、Agent 工具调用和 Docker 部署串成了一条完整工程链路。如果你正在找一个能写进简历、能讲清架构、能覆盖 SpringAI / SpringAIAlibaba学习、技术点的项目，DD_Rag 值得 Star。

Overview: DD_Rag 是一个基于 SpringAI 的 RAG 与 AI 助手项目，旨在提供一个完整的工程链路，包括权限隔离、文档入库、混合检索、证据约束问答、会话式 Agent 和模型治理。这个项目是 SpringAI / SpringAIAlibaba 学习和技术点的绝佳项目，适合写进简历和讲清楚架构。 Key features: - 组织知识库：管理员和普通用户角色分离。 - 文档全生命周期管理：上传、解析、清洗、切片、异入库、状态追踪与删除。 - 混合检索问答：使用 PgVector、Elasticsearch 和 RRF 融合技术。 - Assistant 会话式功能：多轮会话、流式响应、短期记忆摘要和 KB_SEARCH 工具调用。 - 模型平台：平台共享模型、BYOK、场景路由、调用账本和模型治理。 - 调用治理：会话互斥、并发限制、超时与取消。 Tech stack: Java 21、Spring Boot、Spring AI、Spring AI Alibaba、MyBatis-Plus、Flyway、Vue 3、TypeScript、Vite、Pinia、PostgreSQL + pgvector、Elasticsearch + IK、MinIO、Ollama。 Why it matters: 作为一个面向组织知识库的全栈 RAG 与 AI 助手项目，DD_Rag 提供了权限隔离、文档生命周期管理、混合检索问答、会话

TypeScript 213 stars

yoloshii/ClawMem

On-device memory layer for AI agents. Claude Code, OpenClaw and Hermes. Hooks + MCP server + hybrid RAG search.

### Overview: ClawMem is an on-device memory layer designed for AI agents, including Claude Code, OpenClaw, and Hermes, providing a retrieval-augmented search, hooks, and an MCP server in a single local system. ### Key Features: - **Context-Surfacers**: Surface relevant context on every prompt. - **Bootstrapping Sessions**: Capture and integrate decisions, preferences, milestones, and problems from session transcripts. - **Importing Conversations**: Import conversation exports from Claude Code, ChatGPT, Claude.ai, Slack, and plain text. - **Session Handoffs**: Generate handoffs at session end, ensuring the next session picks up where you left off. - **Decision Learning**: Boost referenced notes and decay unused ones through a feedback loop. - **Query Intent Classification**: Classify query intent (WHY / WHEN / ENTITY / WHAT) to weight search strategies. - **Graph Traversal**: Traverse multi-graphs (semantic, temporal, causal) via adaptive beam search. - **Memory Evolving**: Evolve memory metadata as new documents create or refine connections. - **Causal Inference**: Infer causal relationships between facts extracted from session observations. - **Contradiction Detection**: Detect and auto-decay superseded decisions, with a judge configured for conflict resolution. - **Context Guarding**: Guard against prompt injection in surfaced content. - **Dual-Threshold Merge Safety**: Prevent context bleed in derived insights with name-aware dual-threshold merge safety. ### Tech Stack: TypeScript on Bun. ### Why It Matters: Cl

Python 213 stars

yubol-bobo/Awesome-Multi-Turn-LLMs

This is the official GitHub repository for our survey paper "Beyond Single-Turn: A Survey on Multi-Turn Interactions with Large Language Models".

### Overview: This repository compiles a curated list of papers, datasets, and code repositories focused on multi-turn interactions with large language models. The survey paper, published in Transactions on Machine Learning Research (TMLR) 2026, provides a comprehensive overview of multi-turn LLMs, including task types, improvement methods, and open challenges. The repository aims to provide a comprehensive resource for researchers and practitioners in the field. ### Key Features: - **Instruction Following Tasks:** - Instruction Following in General (Mixed) - Instruction Following in Math - Instruction Following in Coding - Instruction Following in Discussion - **Conversational Engagement Tasks:** - Conversational Engagement in General (Mixed) - Conversational Engagement in Roleplay - Conversational Engagement in Healthcare - Conversational Engagement in Education - Conversational Engagement in Jailbreak ### Tech Stack: Python ### Why it Matters: This repository serves as a valuable resource for researchers and practitioners in the field of multi-turn interactions with large language models, providing a comprehensive overview of the field and a curated collection of relevant papers, datasets, and code repositories.

Python 213 stars

zjunlp/IEPile

[ACL 2024] IEPile: A Large-Scale Information Extraction Corpus

Overview: IEPile is a large-scale information extraction corpus, released at ACL 2024, containing 0.32B tokens of high-quality bilingual (Chinese and English) instruction data for training models. Key features: - **Large-scale dataset**: Contains 0.32B tokens of high-quality bilingual (Chinese and English) instruction data. - **Schema-based information extraction**: Designed for schema-based information extraction tasks. - **Bilingual support**: Available in both Chinese and English. - **Model training**: Ideal for fine-tuning and evaluation of information extraction models. Tech stack: Python Why it matters: IEPile provides a comprehensive and high-quality dataset for researchers and developers working on information extraction tasks, particularly those focused on schema-based information extraction. This dataset is essential for training and evaluating models that require bilingual support and large-scale data. Repository: zjunlp/IEPile Primary language: Python GitHub stars: 213 GitHub description: [ACL 2024] IEPile: A Large-Scale Information Extraction Corpus

TypeScript 212 stars

ChetanXpro/nodejs-whisper

NodeJS Bindings for Whisper - the CPU version of OpenAI's Whisper, as initially crafted in C++ by ggerganov.

Overview: This project provides Node.js bindings for OpenAI's Whisper model, enabling audio-to-text transcription. It supports various output formats and optimizes for CPU performance. Key features: - Automatically converts audio to 16 kHz WAV format - Outputs transcripts in multiple formats (.txt, .srt, .vtt, .json, .wts, .lrc) - Supports timestamp precision to single word - Option to split on word rather than token - Optional translation from source language to English - Converts audio format to WAV for whisper compatibility Tech stack: TypeScript Why it matters: This project simplifies the process of using OpenAI's Whisper model for transcription, making it accessible for developers and automating the conversion of audio to text efficiently.

Jupyter Notebook 212 stars

ForeverBlue816/GRACE

[ICML 2026] GRACE: deployable INT4 VLM via quantization-aware distillation.

### Overview: GRACE-VLM is a deployable INT4 Vision-Language Model (VLM) developed for the 2026 International Conference on Machine Learning (ICML). It combines confidence-aware knowledge distillation with group-wise quantization-aware training to achieve high performance with real packed INT4 weights. ### Key Features: - Deployable INT4 Vision-Language Models via Quantization-Aware Distillation. - Scores 76.7 across seven VLM benchmarks, compared to 76.3 for the 8B teacher. - Retains 98% of the GRACE BF16 average for the W4G128 student. - Real packed INT4 weights through the GRACE AWQ loader. - Standard Transformers checkpoint for fine-tuning. ### Tech Stack: Jupyter Notebook ### Why It Matters: GRACE-VLM offers a significant improvement in performance while maintaining high accuracy, making it a valuable tool for deploying large language models in resource-constrained environments.

Python 212 stars

Paparusi/legal-ai-agent

Overview: This project is an AI legal assistant designed for Vietnamese businesses, offering features like legal research, contract review, compliance checks, clause drafting, contract summaries, comparison, and full document control. Key features: - Legal search across 40,000+ Vietnamese legal documents - Contract review with risk analysis and amendment suggestions - Compliance check against Vietnamese law - Clause drafting for confidentiality, penalty, termination, and force majeure clauses - Contract summary and comparison tools - Company memory across chat sessions - Full document control with autonomous editing, reading, writing, and deletion capabilities Tech stack: Python, FastAPI, Claude (AI model) Why it matters: This project aims to provide a user-friendly interface for businesses to manage their legal documents and contracts efficiently, leveraging AI to automate tasks and improve accuracy in legal research and compliance checks.

Python 212 stars

huggingface/optimum-habana

Easy and lightning fast training of 🤗 Transformers on Habana Gaudi processor (HPU)

Overview: Optimum for Intel Gaudi is a Python interface that simplifies the training and inference of Hugging Face Transformers and Diffusers models on Intel Gaudi AI Accelerators. Key features: - Easy model loading and inference - Compatibility with different downstream tasks - Officially validated models and tasks Tech stack: Python Why it matters: Optimum for Intel Gaudi streamlines the training and inference process on Intel Gaudi accelerators, making it easier and faster for developers to deploy state-of-the-art models.

Python 212 stars

yrcong/flatten

Pytorch Implementation of FLATTEN: optical FLow-guided ATTENtion for consistent text-to-video editing (ICLR 2024)

Overview: FLATTEN is a Pytorch implementation of a novel method for text-to-video editing, specifically designed to address the challenge of ensuring visual consistency in edited videos. This approach leverages optical flow to guide attention in the attention module of a diffusion model, thereby improving the quality of the edited video. Key features: - Optical flow-guided attention module - Ensures visual consistency in text-to-video editing - Training-free and easily integrable into existing diffusion-based text-to-video editing methods Tech stack: Python Why it matters: FLATTEN provides a unique solution to the challenge of maintaining visual consistency in text-to-video editing, making it a valuable addition to the field of video editing and AI-driven content creation. The method's ability to enforce attention across patches on the same flow path across frames is particularly noteworthy, as it addresses a significant challenge in the field. Repository: yrcong/flatten Primary language: Python GitHub stars: 212 GitHub description: Pytorch Implementation of "FLATTEN: optical FLow-guided ATTENtion for consistent text-to-video editing". The paper has been accepted at ICLR 2024. The repository includes instructions for downloading Stable Diffusion 2.1, installing necessary packages, and running the script for text-to-video editing. The README provides detailed information on the requirements, usage, and editing tricks, including tips on using negative prompts and increasing classifier-free guidance scale. The project is accompanied by a teaser video and a BibTeX entry for academic reference.

C++ 211 stars

AlayaDB-AI/AlayaLite

AlayaLite – A Fast, Flexible Vector Database for Everyone.

### Overview: AlayaLite is a fast, flexible vector database designed for everyone, offering high performance, elastic scalability, and easy customization. It seamlessly integrates modern vector techniques into a well-designed architecture, making it accessible and efficient for a wide range of applications. ### Key Features: - **High Performance**: Modern vector techniques integrated into a well-designed architecture. - **Elastic Scalability**: Seamlessly scale across multiple threads, optimized by C++20 coroutines. - **Adaptive Flexibility**: Easy customization for quantization methods, metrics, and data types. - **Two index paths in one package**: an in-memory graph + RaBitQ path for low-latency retrieval, and the **LASER** on-disk Quantized Graph index for billion-scale workloads that do not fit in RAM. - **Ease of Use**: Intuitive APIs in Python. ### Tech Stack: C++ ### Why it Matters: AlayaLite is a powerful tool for vector database applications, enabling seamless integration of modern vector techniques with efficient scalability and customization options, making it ideal for a wide range of applications. ### Repository: AlayaDB-AI/AlayaLite ### GitHub Stars: 210 ### Primary Language: C++ ### Documentation: - [Client User Guide](https://github.com/AlayaDB-AI/AlayaLite/blob/main/docs/CLIENT_USER_MANUAL.md) - [LASER Guide](https://github.com/AlayaDB-AI/AlayaLite/blob/main/docs/LASER.md) ### Quick Start: - **Client User

Java 211 stars

El-12stu/RelaMind

基于 AI 的个人成长轨迹分析系统，通过记录生活、回顾历史、分析模式帮助用户更好地理解自己，实现持续成长。包含智能路由、RAG 检索、自主任务智能体等功能。

### Overview: RelaMind 是一个基于人工智能的个人成长轨迹分析系统，旨在通过记录生活、回顾历史和分析模式帮助用户更好地理解自己，实现持续成长。 ### Key features: - **智能对话**：自动识别用户意图，提供情感支持和引导记录。 - **成长日记**：记录和管理个人成长，支持多种心情选择和标签管理。 - **Manus 任务助手**：基于 ReAct 模式，支持多步骤任务规划和文件生成。 - **敏感词检测**：内置敏感词检测功能，保护系统安全。 ### Tech stack: Java ### Why it matters: 通过智能对话、成长日记和任务助手等功能，RelaMind 提供了全面的个人成长支持，帮助用户更好地理解自己和实现持续成长。

Python 211 stars

MiniMax-AI/SynLogic

[NeurIPS 2025] The official repo of SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond

Overview: SynLogic is a comprehensive logical reasoning data synthesis framework that generates diverse, verifiable reasoning data at scale. It addresses the gap in high-quality logical reasoning training data for developing general reasoning capabilities in Large Language Models. Key features: - Covers 35 diverse logical reasoning tasks - Enables controllable difficulty synthesis - Provides verifiable rewards through rule-based verification - Achieves SOTA performance among open-source datasets Tech stack: Python Why it matters: SynLogic is crucial for training models to improve logical reasoning capabilities, making it essential for advancing the field of AI and language models.

Python 211 stars

Renumics/renumics-rag

Visualization for a Retrieval-Augmented Generation (RAG) Assistant 🤖❤️📚

### Overview: Renumics RAG: A Visualization Tool for Retrieval-Augmented Generation (RAG) Assistant ### Key features: - Interactive visualization of RAG data - Supports GPU and CPU environments - Integrates LangChain and Streamlit for seamless integration ### Tech stack: Python ### Why it matters: Renumics RAG is a powerful tool for developers and researchers looking to enhance their RAG-based applications with interactive visualization capabilities. The integration of LangChain and Streamlit ensures a smooth and user-friendly experience, making it ideal for both local and cloud-based deployments. The support for GPU and CPU environments ensures optimal performance, making it suitable for a wide range of applications.

Unknown 211 stars

YuanchenBei/Awesome-Graphs-Meet-Agents

[Up-to-date] A curated list of resources on graph-empowered agents and agent-facilitated graph learning (Graphs Meet Agents & Agentic Graph Engineering).

### Overview: This repository is a curated list of resources related to graph-empowered agents and agentic graph engineering. It is based on the survey paper "[Graphs Meet AI Agents: Taxonomy, Progress, and Future Opportunities](https://arxiv.org/pdf/2506.18019)". ### Key features: - Knowledge Graph Engineering for Multi-Agentic Systems: The Anthropic Playbook - Graph Engineering - Task Reasoning - (NAACL 2021) QA-GNN: Reasoning with language models and knowledge graphs for question answering - (ICLR 2024) Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph - (ICLR 2024) Reasoning on graphs: Faithful and interpretable large language model reasoning - (IJCAI 2024) Kg-cot: Chain-of-thought prompting of large language models over knowledge graphs for knowledge-aware question answering - (ACL 2024) Mindmap: Knowledge graph prompting sparks graph of thoughts in large language models - (WWW 2025) Paths-over-graph: Knowledge graph empowered large language model reasoning - (ICML 2025) GIVE: Structured Reasoning of Large Language Models with Knowledge-Graph-Inspired Veracity Extrapolation - Graph for Agent Planning - Task Reasoning - (NAACL 2021) QA-GNN:

Unknown 211 stars

choyingw/Awesome-Monocular-Depth

A curated list of recent monocular depth estimation papers

**Overview:** This repository is a curated list of recent monocular depth estimation papers, focusing on recent work after 2020. It aims to provide a comprehensive overview of the latest advancements in the field, inspired by the awesome-computer-vision repository. **Key features:** - High Performance: Papers that offer high performance metrics for monocular depth estimation. - Meta-learning: Papers that use meta-learning techniques for improving generalizability. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Zero-shot Evaluation: Papers that perform well on zero-shot evaluation scenarios. - High-Resolution: Papers that achieve high-resolution results for monocular depth estimation. - Universal Models: Papers that offer universal models for monocular depth estimation. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Fast Estimation: Papers that provide fast monocular depth estimation methods. - Zero-shot Performance: Papers that match the performance of depth-anything on zero-shot evaluation. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Versatile Models: Papers that focus on developing versatile models for monocular depth estimation. - Versatile Models:

TypeScript 211 stars

eren23/openflipbook

Open-source flipbook.page clone — every page is an AI-generated illustration, click anywhere to explore deeper. Next.js + FastAPI + Modal. BYO keys.

### Overview: Explore illustrated worlds with openflipbook, a self-hosted and MIT-licensed flipbook clone. Every page is an AI-generated illustration, and you can click anywhere to navigate deeper. Built with Next.js and FastAPI, it's BYO keys, inspired by flipbook.page. ### Key Features: - **Illustrated Pages**: Every page is an AI-generated illustration. - **Click-Through Navigation**: Click anywhere to explore deeper. - **Self-Hosted**: Open-source and self-hosted, inspired by [flipbook.page](https://flipbook.page). ### Tech Stack: - Next.js - FastAPI - Modal ### Why It Matters: This open-source project offers a similar experience to flipbook.page, but with the flexibility to use your own keys, storage, and backend. It's a fun and interactive way to explore and edit illustrated worlds. ### Repository: The repository is maintained by Eren23 and is hosted on GitHub. It has received 210 GitHub stars, indicating its popularity and usefulness. ### Primary Language: The primary language used in the project is TypeScript. ### GitHub Stars: The repository has received 210 GitHub stars, indicating its popularity and usefulness. ### Demo: See the demo in the README for a visual representation of how the app works. The app is self-hosted and inspired by [flipbook.page](https://flipbook.page). ### Why This Exists: flipbook.page is fun but closed; this project aims to provide the same loop on a stack you own, using your own keys,

Python 211 stars

francozanardi/pictex

A Python library for efficient image generation using CSS Flexbox

Overview: PicTex is a Python library designed for generating high-quality images using CSS Flexbox. It simplifies the process of creating complex layouts and renders them efficiently, making it perfect for various applications such as Open Graph, social media graphics, and automated report generation. Key features: - Built on industry-standard CSS Flexbox principles with near-complete property support. - Composes complex visuals by nesting powerful layout primitives like `Row`, `Column`, and `Image`. - Supports gradients, multiple shadows, borders with rounded corners, and text decorations. - Allows for custom fonts, variable fonts, line height, alignment, and professional text shaping using HarfBuzz for kerning, ligatures, and complex scripts. - Seamlessly renders emojis and multilingual text. - Automatically handles BiDi text rendering. - Supports flexible output options: raster (PNG/JPEG/WebP) and vector (SVG) formats. - Powered by Google's Skia graphics engine for high-quality rendering. Tech stack: PicTex is built on Python, utilizing Skia, Taffy, and HarfBuzz for its layout and rendering capabilities. Why it matters: PicTex simplifies the process of creating visually appealing and high-quality images without the need for complex JavaScript or other external libraries. It leverages the power of CSS Flexbox, making it easy to create complex layouts that can be easily converted into various formats for different use cases.

JavaScript 211 stars

kevin-hs-sohn/hipocampus

Drop-in memory harness for AI agents — 3-tier memory, compaction tree, hybrid search. One command to set up. Works with Claude Code and OpenClaw.

### Overview: Hipocampus is a proactive memory harness designed for AI agents, providing a zero-configuration solution that compacts and indexes past interactions to enable quick retrieval of relevant context. The system is optimized for Claude Code, OpenCode, and OpenClaw, offering a 21.6x improvement over no memory and a 5.1x improvement over BM25 search alone. ### Key Features: - **Compaction Tree:** A hierarchical indexing system that compacts and indexes past interactions. - **Hybrid Search:** A compaction tree combined with vector search for more comprehensive context retrieval. - **Easy Setup:** A single command to set up and integrate with Claude Code, OpenCode, and OpenClaw. ### Tech Stack: - JavaScript ### Why It Matters: Hipocampus addresses the challenge of AI agents forgetting context, especially in complex, multi-session interactions. By providing a holistic view of past knowledge, it ensures that relevant information is always accessible, preventing retrieval failures and improving the overall performance and reliability of the AI agent.

Go 210 stars

amikos-tech/chroma-go

The Go client for Chroma vector database

### Overview: Chroma Go is a Go client library for interacting with the Chroma vector database, providing a simple and efficient way to manage and query vector data. Designed for developers, it simplifies the integration of Chroma into Go projects, making it easier to build applications that leverage vector databases. ### Key Features: - **Vector Database Integration**: Provides a straightforward interface for interacting with Chroma's vector database. - **Compatibility**: Supports both `v0.3.x` and `v0.4.x` releases of Chroma, ensuring compatibility with various versions of Chroma. - **Persistent Client**: Automatically handles the downloading and caching of shim libraries, making it easy to start a local instance without external dependencies. - **Self-Hosted (HTTP)**: Allows connecting to a Chroma server running on a local HTTP endpoint, enabling remote access to vector data. ### Tech Stack: Go ### Why It Matters: Chroma Go simplifies the process of integrating Chroma into Go projects, making it easier to build applications that leverage vector databases. The library is designed with developer convenience in mind, reducing the complexity of setting up and using vector databases in Go applications.