ArchestackGitHub ↗ (opens in a new tab)

Open-source alternatives to Glean

RAGFlow leads with 90, ahead of deeptutor (89) and LightRAG (89). Updated from public GitHub data.

The highest-scoring open-source projects in RAG, knowledge and search. They are picks, not exact replacements, so read the weaknesses before you switch. The same picks fit if you are leaving Glean AI.

Open-source alternatives to Glean
RankProjectScore
1RAGFlowRAG engine with deep document parsing, template chunking and agentic retrieval Live demo ↗ (opens in a new tab)RAG, knowledge and search92.8k starsApache-2.090 out of 100
2deeptutorSelf-hosted AI tutor with knowledge bases, guided learning, books and quizzesRAG, knowledge and search42.4k starsApache-2.089 out of 100
3LightRAGGraph-based RAG framework with a REST API server and web UIRAG, knowledge and search40.3k starsMIT89 out of 100
#4MaxKBEnterprise knowledge-base agent platform with RAG, workflows and MCP toolsRAG, knowledge and search23k starsGPL-3.086 out of 100
#5Open NotebookSelf-hosted NotebookLM alternative with podcasts and 20+ model providersRAG, knowledge and search40.5k starsMIT83 out of 100
#6pageindexVectorless RAG that builds a tree index of a document and searches it with an LLM Live demo ↗ (opens in a new tab)RAG, knowledge and search39.6k starsMIT82 out of 100
#7SurfSenseDesktop NotebookLM-style agent that researches and edits local documentsRAG, knowledge and search16.6k starsApache-2.082 out of 100
#8GPT ResearcherResearch agent that writes cited reports from web and local documentsRAG, knowledge and search30.3k starsApache-2.081 out of 100
#9WeKnoraKnowledge base framework with RAG Q&A, tool-using agents and auto-generated wikisRAG, knowledge and search35.3k starscustom license80 out of 100
#10graphragPipeline that builds knowledge graphs from text for LLM question answeringRAG, knowledge and search36.4k starsMIT78 out of 100

Reviews

190 out of 100

RAGFlow

RAG engine with deep document parsing, template chunking and agentic retrieval

Live demo ↗ (opens in a new tab)92.8k starsApache-2.0last commit Oct 2026

RAGFlow parses complex documents (PDF, Word, Excel, PowerPoint, scans, web pages) with its DeepDoc layout, OCR and table recognition, then chunks them by template and answers with traceable citations. It adds multi-step agentic retrieval with four thinking modes, plus knowledge compilation into wikis, graphs and mind maps. Services are written in Go and deployed with Docker Compose, with a web UI and HTTP API.

Strengths

  • Template-based chunking with a visual view of chunks so humans can correct them
  • Answers include traceable citations back to source chunks
  • Ingests many formats, plus sitemap crawling and Google BigQuery sync
  • Configurable LLM, embedding and reranker models; multiple recall with fused re-ranking

Weaknesses

  • Heavy stack: MySQL, MinIO, NATS, Kvrocks and a document engine; 16 GB RAM suggested
  • macOS is not supported for the Docker deployment
  • DeepDoc runs on CPU inference in the 1.0 release, which may slow parsing
  • Current release is 1.0.0-rc1, a release candidate rather than a stable 1.0
  • RAM ≥ 16 GB
  • no GPU
  • Docker + Compose
  • Needs Docker, Docker Compose, MySQL, Elasticsearch or Infinity, MinIO, NATS, Kvrocks
  • Models: Configurable LLM, embedding and reranker models via provider page
  • port 80
Live demo ↗ (opens in a new tab)Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
289 out of 100

deeptutor

Self-hosted AI tutor with knowledge bases, guided learning, books and quizzes

42.4k starsApache-2.0last commit Oct 2026

DeepTutor is a Python backend with a Next.js 16 web UI for studying from your own documents. It combines chat over knowledge bases (LightRAG, PageIndex, GraphRAG, LlamaIndex, FAISS) with guided learning, generated books, quizzes, question banks and deep research. It also offers multi-user accounts, IM-connected Partners, MCP services and a CLI.

Strengths

  • Several retrieval engines selectable per knowledge base: LightRAG, PageIndex, GraphRAG, LlamaIndex
  • Covers study workflows: guided learning, mastery gates, quizzes, question bank, daily practice
  • Multi-user deployment with per-user isolation, admin grants and learner/guardian accounts
  • Wide provider and parser coverage, including Codex OAuth, MinerU, Tika and PyMuPDF4LLM

Weaknesses

  • Fast release cadence with breaking refactors (v1.6.3) makes upgrades risky
  • Large surface area: Partners, MCP, CLI apps and plugins add configuration overhead
  • RAM, GPU and port requirements are not stated in the visible README
  • Aimed at learning and study; less suited to general chat or enterprise search
  • Docker + Compose
  • Needs Python 3.11+, Node.js (Next.js 16), Docker (optional)
  • Models: OpenAI-compatible providers, Anthropic, Gemini, OpenAI Codex OAuth, Lemonade (local)
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
389 out of 100

LightRAG

Graph-based RAG framework with a REST API server and web UI

40.3k starsMITlast commit Oct 2026

LightRAG indexes documents into a knowledge graph plus vector embeddings and answers queries with dual-level (entity and concept) retrieval. It ships as a Python library and as lightrag-server, which provides a REST API and a web UI for inserting, querying and visualizing the graph. It supports incremental updates, document deletion with graph rebuild, reranking, and several parsing and chunking options.

Strengths

  • Document deletion rebuilds affected graph entries from the LLM cache; inserts are incremental
  • Storage backends include PostgreSQL, Neo4j, MongoDB and OpenSearch, plus local file stores
  • Parsers: MinerU, Docling and a native engine; four chunking strategies
  • Setup wizard generates .env and compose files; RAGAS evaluation and Langfuse tracing supported

Weaknesses

  • Default in-memory/file stores are stated as not for production
  • Server binds to 0.0.0.0 and endpoints are public until auth is set in .env
  • Graph extraction needs LLM calls at index time; RAM and GPU needs are unstated
  • Optional features need extra steps: spaCy models for smart_heading, libcairo for SVG
  • Docker + Compose
  • Needs LLM provider, embedding model, PostgreSQL (recommended for production), Neo4j (optional), MongoDB (optional), OpenSearch (optional), MinerU or Docling (optional)
  • Models: LLMs configured via .env, open-source LLMs such as Qwen3-30B-A3B, rerankers
Documentation ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#486 out of 100

MaxKB

Enterprise knowledge-base agent platform with RAG, workflows and MCP tools

23k starsGPL-3.0last commit Oct 2026

MaxKB runs as one Docker container (port 8080, state in one volume) with a RAG pipeline that uploads or crawls documents, a workflow engine with function library and MCP tool use, and zero-code embedding into other systems. It works with private models (DeepSeek, Llama, Qwen) and public APIs (OpenAI, Claude, Gemini, MiniMax) and handles text, image, audio and video. Built on Django, LangChain and PostgreSQL.

Strengths

  • Single docker run with all data under one mounted volume
  • Workflow engine with function library and MCP tool calling
  • Crawls online documents into the knowledge base automatically
  • Multimodal input and output: text, image, audio, video

Weaknesses

  • GPL-3.0 limits bundling into proprietary products
  • Ships with default admin password MaxKB@123..
  • README gives no hardware guidance or provider configuration detail
  • Detailed docs are on maxkb.cn, partly in Chinese
  • no GPU
  • Models: OpenAI, Claude, Gemini, MiniMax, DeepSeek, Llama, Qwen as private models
  • port 8080
Repository on GitHub ↗ (opens in a new tab)
#583 out of 100

Open Notebook

Self-hosted NotebookLM alternative with podcasts and 20+ model providers

40.5k starsMITlast commit Oct 2026

Open Notebook collects PDFs, audio, video, web pages and Office files into notebooks and offers cited chat, full-text and vector search, notes and multi-speaker podcast generation. It runs as two containers (SurrealDB plus a FastAPI/Next.js app) and talks to OpenAI, Anthropic, Google, Mistral, Groq, Ollama, LM Studio or any OpenAI-compatible server. A REST API and MCP integration expose the same features.

Strengths

  • 20+ providers, including Ollama and LM Studio for fully local runs
  • Podcasts with 1 to 4 speakers and custom episode profiles
  • REST API on port 5055 and an MCP server for Claude Desktop or VS Code
  • Two-service Docker Compose; keys stored encrypted with OPEN_NOTEBOOK_ENCRYPTION_KEY

Weaknesses

  • Single-user; multi-user support is only a future direction in VISION.md
  • No password by default and ports 8502/5055 bind to all interfaces
  • Anthropic and Groq offer no embeddings, so a second provider is needed
  • UI in 14 languages but provider setup is manual per model type
  • no GPU
  • Docker + Compose
  • Needs SurrealDB
  • Models: OpenAI, Anthropic, Google, Mistral, Groq, DeepSeek, xAI, OpenRouter, Cohere, Ollama, LM Studio, oMLX and any OpenAI-compatible endpoint
  • port 8502
  • README: alternative to NotebookLM
Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#682 out of 100

pageindex

Vectorless RAG that builds a tree index of a document and searches it with an LLM

Live demo ↗ (opens in a new tab)39.6k starsMITlast commit Oct 2026

PageIndex builds a hierarchical tree index from each long PDF and has an LLM reason its way through that tree to answer questions, with no vector database and no chunking. The Python SDK runs in local mode with your own LLM API key, or against PageIndex Cloud, which adds OCR, image understanding and an MCP server. Local mode handles text-based PDFs only and cites at page level.

Strengths

  • No vector DB or chunking; retrieval is traceable to tree nodes and page references
  • Local indexing costs about $0.001 per page; 9 to 1,098 pages took 13 seconds to 4.5 minutes
  • Reports 98.7% on FinanceBench; the repo documents a separate OSS benchmark with a runner
  • Same client switches between local and cloud; supports agent SDK integration

Weaknesses

  • Local mode handles text-based PDFs only; OCR and image understanding require the paid Cloud
  • Metadata, folders, MCP server and File System for large corpora are Cloud-only
  • Query accuracy depends on a strong chat model, which raises per-question cost
  • Needs an external LLM API for indexing and chat; no fully offline setup is described
  • no GPU
  • Needs LLM API key (OpenAI or other supported provider), PageIndex API key (Cloud mode only)
  • Models: OpenAI models, Claude via Claude Code subscription, PageIndex Cloud compatible models
Live demo ↗ (opens in a new tab)Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#782 out of 100

SurfSense

Desktop NotebookLM-style agent that researches and edits local documents

16.6k starsApache-2.0last commit Oct 2026

SurfSense is a desktop app for Windows, macOS (Apple Silicon) and Linux that answers questions over your files with citations and can also create, edit and convert PDF, Word, Excel and PowerPoint documents. The index stays on disk and there is no account. You choose the model: any OpenAI-compatible API or a local one. Edits always produce a new copy.

Strengths

  • Index stays on your disk; no account required
  • Edits PDF, Word and Excel files as copies, never altering originals
  • Transcribes audio locally and reads scanned pages via OCR
  • Works with any OpenAI-compatible API or a local model

Weaknesses

  • Built-in plugins (Google Search, Web Crawl, etc.) need a SurfSense license; trial is 30 days
  • Word and PowerPoint to PDF needs LibreOffice, not included in the installer
  • Image generation needs a model you configure; none is chosen by default
  • Hosted web app is being sunset; only the desktop app remains
  • GPU optional
  • Docker + Compose
  • Needs LibreOffice (optional, for Office conversion)
  • Models: OpenAI-compatible APIs, local models
  • README: alternative to NotebookLM
Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#881 out of 100

GPT Researcher

Research agent that writes cited reports from web and local documents

30.3k starsApache-2.0last commit Oct 2026

GPT Researcher splits a query into research questions with a planner agent, scrapes sources in parallel, filters the context and writes a cited report. It works over web search (Tavily by default) and local files such as PDF, CSV, Excel, Markdown, PowerPoint and Word. It runs as a FastAPI server with a static or NextJS frontend, or as the gpt-researcher pip package.

Strengths

  • Covers both web research and local documents (PDF, CSV, Excel, Markdown, PowerPoint, Word)
  • Usable as a pip package, a FastAPI server with two frontends, or via Docker Compose
  • Supports MCP data sources, recursive deep research, and LangGraph or AG2 multi-agent workflows
  • Context filter falls back to local BM25 keyword ranking when no TypeSafe key is set

Weaknesses

  • Needs an LLM API key and a search API key (OpenAI and Tavily in the README examples)
  • Default Jev context filter requires a TypeSafe API key and a hosted call
  • Deep research is quoted at about 5 minutes and $0.4 per run with o3-mini
  • README calls it experimental and provided as-is, with no warranty
  • no GPU
  • Docker + Compose
  • Needs OpenAI API key, Tavily API key, Python 3.12+, TypeSafe API key (optional, Jev filter), Google API key (optional, image generation)
  • Models: OpenAI, OpenAI-compatible APIs via OPENAI_BASE_URL, Google Gemini (inline images)
  • port 8000
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#980 out of 100

WeKnora

Knowledge base framework with RAG Q&A, tool-using agents and auto-generated wikis

35.3k starscustom licenselast commit Oct 2026

WeKnora is a Go-based knowledge framework from Tencent that parses documents (PDF, Word, Excel, images, XMind and others) into knowledge bases. It then serves hybrid-search RAG with citations, a multi-step agent with sandboxed skills and MCP tools, and an auto-generated wiki with a knowledge graph. It syncs from sources like Feishu, Confluence, Notion and GitLab, answers in IM apps such as WeCom, Slack and Telegram, and can also run as a built-in MCP server.

Strengths

  • Many data source connectors: Feishu, Confluence, Notion, GitLab, Yuque, DingTalk, Seafile, RSS
  • Multi-workspace RBAC with four roles, audit log and scoped API keys
  • Swappable LLMs, vector databases and storage; 29 built-in model vendors including Ollama
  • Lite single-binary mode (SQLite, in-memory queue) alongside Docker Compose and Helm chart

Weaknesses

  • License reported as NOASSERTION by detection, while the README badge says MIT
  • Full feature set needs several services (Neo4j, MinIO, Langfuse via profiles)
  • Desktop app has no published installer; must be built from source
  • Frequent breaking changes between releases, such as sandbox commands running as root in v0.8.2
  • GPU optional
  • Docker + Compose
  • Needs Docker, Docker Compose, Neo4j (optional), MinIO (optional), Langfuse (optional), Ollama (optional)
  • Models: OpenAI, DeepSeek, Qwen, Zhipu, Hunyuan
  • port 80
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#1078 out of 100

graphrag

Pipeline that builds knowledge graphs from text for LLM question answering

36.4k starsMITlast commit Oct 2026

GraphRAG is a Python data pipeline from Microsoft Research that uses an LLM to extract structured data from unstructured text and form a knowledge graph. That graph is used to build context for question answering over private data. It runs from a command line (`graphrag init`, indexing, querying) with configurable settings and tunable prompts. The project is in maintenance mode: bug fixes and dependency updates only.

Strengths

  • Backed by a published paper (arXiv 2404.16130) and a Microsoft Research blog post
  • Prompt tuning guide for adapting extraction to your own data
  • Installable from PyPI, with a CLI quickstart and hosted documentation
  • MIT license, with a release in September 2026

Weaknesses

  • Maintenance mode: no new features and no new PRs accepted
  • Indexing can be expensive in LLM calls; README advises starting small
  • Not an officially supported Microsoft offering; code is a demonstration
  • Config format changes between versions require re-running `graphrag init --force`
  • Needs LLM provider
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)