ArchestackGitHub ↗ (opens in a new tab)

Open-source alternatives to Glean

The highest-scoring open-source projects in RAG and knowledge and Search. They are picks, not exact replacements, so read the weaknesses before you switch.

Open-source alternatives to Glean
RankProjectScore
1LightRAGGraph-plus-vector RAG server with web UI and Ollama-compatible APIRAG and knowledge, 40.1k stars, MIT76 out of 100
2Open NotebookSelf-hosted NotebookLM alternative with podcasts and 20+ model providersRAG and knowledge, 40k stars, MIT75 out of 100
3RAGFlowRAG engine with document parsing, agentic retrieval and citations Live demo ↗ (opens in a new tab)RAG and knowledge, 92k stars, Apache-2.072 out of 100
#4MorphicSearch engine that answers with citations and renders rich inline componentsSearch, 9.2k stars, Apache-2.070 out of 100
#5WeKnoraEnterprise knowledge base combining RAG Q&A, agents and generated wikisRAG and knowledge, 33k stars, custom license68 out of 100
#6SurfSenseOffline NotebookLM alternative that turns documents into decks, reports and podcastsRAG and knowledge, 16.3k stars, custom license68 out of 100
#7MaxKBEnterprise knowledge-base agent platform with RAG, workflows and MCP toolsRAG and knowledge, 22.9k stars, GPL-3.066 out of 100
#8Local Deep ResearchAgentic research assistant with local LLMs, SearXNG and encrypted librariesSearch, 9.2k stars, MIT65 out of 100
#9GPT ResearcherResearch agent that writes cited reports from web and local documentsSearch, 30k stars, Apache-2.064 out of 100
#10PrivateGPTAnthropic-style API layer for private RAG on local inference serversRAG and knowledge, 57.6k stars, Apache-2.063 out of 100

Reviews

176 out of 100

LightRAG

Graph-plus-vector RAG server with web UI and Ollama-compatible API

40.1k stars, MIT, last commit Sep 2026

LightRAG indexes documents into a knowledge graph plus vector store and queries both layers, as a lighter alternative to Microsoft GraphRAG. The server package ships a REST API, a web UI for inserting and visualizing the graph, and Ollama-compatible /api routes for chat frontends. Parsing runs via MinerU, Docling or a native engine; production storage goes to PostgreSQL, Neo4j, MongoDB, Milvus or OpenSearch.

Strengths

  • Dual-level graph and vector retrieval with fewer LLM calls than community-report GraphRAG
  • Incremental updates and document deletion with graph regeneration from the LLM cache
  • Three parsing engines and four chunking strategies, including paragraph-semantic
  • Separate LLM settings per role: extract, query, keywords and VLM

Weaknesses

  • Default KV, vector and graph stores are in-memory with file persistence, not for production
  • Server binds 0.0.0.0 with every endpoint public until auth is configured
  • Ollama-compatible /api routes stay open even with auth unless WHITELIST_PATHS is set
  • docx smart headings and SVG rendering need extra spaCy models and libcairo
  • no GPU
  • Docker + Compose
  • Needs PostgreSQL (recommended for production), Neo4j (optional), MongoDB (optional), Milvus (optional), OpenSearch (optional)
  • Models: LLM and embedding providers configured in .env, tested with open models such as Qwen3-30B-A3B
Repository on GitHub ↗ (opens in a new tab)
275 out of 100

Open Notebook

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

40k stars, MIT, last 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)
372 out of 100

RAGFlow

RAG engine with document parsing, agentic retrieval and citations

Live demo ↗ (opens in a new tab)92k stars, Apache-2.0, last commit Oct 2026

RAGFlow parses documents (Word, slides, Excel, TXT, images, scans, web pages) with template-based chunking, then retrieves with multiple recall and fused re-ranking to produce answers with traceable citations. Recent releases add agentic multi-step retrieval with four thinking modes and Knowledge Compilation into wikis, graphs, trees and mind maps. Models are configured by name, address and API key for the LLM, embedding and reranker.

Strengths

  • Chunk visualization lets you inspect and correct parsing before retrieval
  • Citations link answers back to source chunks
  • Ingests sitemaps and Google BigQuery with incremental sync
  • Apache-2.0 with prebuilt Docker Compose deployment

Weaknesses

  • Stack needs MySQL, MinIO, NATS, Kvrocks, ClickHouse and a document engine
  • Go backend not supported on macOS; Linux x86_64 host required
  • DeepDoc OCR and layout analysis run on CPU only in 1.0
  • Current release is 1.0.0-rc1, a release candidate
  • RAM ≥ 16 GB
  • no GPU
  • Docker + Compose
  • Needs Elasticsearch or Infinity, MySQL, MinIO, NATS JetStream, Kvrocks, ClickHouse
  • Models: configurable LLM, embedding, reranker
  • 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)
#470 out of 100

Morphic

Search engine that answers with citations and renders rich inline components

9.2k stars, Apache-2.0, last commit Oct 2026

Morphic runs web searches and returns cited answers, rendering inline components such as images, grids and headings from a streamed JSON spec instead of plain markdown. It offers Quick and Adaptive search modes and works with OpenAI, Anthropic, Google, Ollama, Vercel AI Gateway and OpenAI-compatible models. Docker Compose brings up PostgreSQL, Redis, SearXNG and the app, with chat history stored in PostgreSQL.

Strengths

  • Compose file bundles PostgreSQL, Redis and SearXNG, so no search API key is required
  • Supports Tavily, SearXNG, Brave and Exa as search providers
  • Model selector detects providers, including local Ollama and OpenAI-compatible endpoints
  • Auth is switchable between Supabase, better-auth and none

Weaknesses

  • Full stack needs four containers: PostgreSQL, Redis, SearXNG and the app
  • Supabase is the default auth provider unless ENABLE_AUTH=false or AUTH_PROVIDER is set
  • Needs at least one AI provider API key or a local Ollama setup
  • README gives no RAM or hardware requirements
  • Docker + Compose
  • Needs PostgreSQL, Redis, SearXNG, Supabase Auth (optional)
  • Models: OpenAI, Anthropic, Google, Ollama, Vercel AI Gateway
  • port 3000
Repository on GitHub ↗ (opens in a new tab)
#568 out of 100

WeKnora

Enterprise knowledge base combining RAG Q&A, agents and generated wikis

33k stars, custom license, last commit Oct 2026

WeKnora turns team documents into knowledge bases with three modes: cited RAG answers, an agent that runs multi-step tasks with skills in Docker, E2B or Cube sandboxes, and auto-generated wiki pages with a knowledge graph. It syncs from Feishu, Confluence, GitLab, Notion and RSS, answers in WeCom, Slack and Telegram, and exposes an MCP server. Deploy with Docker Compose, Helm or one Lite binary on SQLite.

Strengths

  • 29 built-in model vendors including OpenAI, DeepSeek, Qwen, Gemini, LiteLLM and Ollama
  • Lite single binary with SQLite and in-memory queue for low-resource hosts
  • Workspace RBAC with four roles, per-resource ownership and audit log
  • Per-workspace MCP endpoints with own token, scope and rate limit

Weaknesses

  • Many integrations target the Chinese ecosystem (WeChat, Feishu, DingTalk, Yuque)
  • Sandbox commands run as root since v0.8.2
  • Maintainers advise against exposing it to the public internet
  • Desktop app has no published installer; hardware requirements live in external docs
  • no GPU
  • Docker + Compose
  • Needs Neo4j (optional profile), MinIO (optional profile), Langfuse (optional profile)
  • Models: OpenAI, DeepSeek, Qwen, Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, LiteLLM, Ollama
  • port 80
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#668 out of 100

SurfSense

Offline NotebookLM alternative that turns documents into decks, reports and podcasts

16.3k stars, custom license, last commit Oct 2026

SurfSense indexes local PDFs, Office files and images into SQLite, answers with citations and turns sources into summaries, flashcards, quizzes, mind maps, editable pptx/docx/xlsx and offline podcasts (Kokoro-82M). It runs a local Qwen3 (six sizes from 0.5 GB) or any OpenAI-compatible API, egress off by default. The supported path is a desktop installer; the Docker stack is community-supported.

Strengths

  • Parser, retrieval model and podcast voice ship in the installer; works with networking off
  • Egress panel off by default, no telemetry or crash reporting
  • Produces editable pptx, docx and xlsx files rather than chat only
  • No account required; keys stored in the OS keychain

Weaknesses

  • Primary product is a desktop app, not a server
  • Self-hosted Docker stack has no SLA and no hosted service behind it
  • Hosted web app retired; export window closes 2026-10-18
  • Plugins and priority support are behind a paid licence; no video overviews
  • no GPU
  • Docker + Compose
  • Models: local Qwen3 in six sizes from 0.5 GB, any OpenAI-compatible API
  • README: alternative to NotebookLM
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#766 out of 100

MaxKB

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

22.9k stars, GPL-3.0, last 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)
#865 out of 100

Local Deep Research

Agentic research assistant with local LLMs, SearXNG and encrypted libraries

9.2k stars, MIT, last commit Oct 2026

Runs multi-step research across the web, academic engines and your own documents using Ollama or any OpenAI-compatible endpoint, with a LangGraph agent that picks engines adaptively and writes cited reports. Each user gets an AES-256 SQLCipher database, and egress scopes limit which engines and providers a run may use. Web UI on port 5000 via Docker, Compose or pip, for privacy-focused researchers.

Strengths

  • Reports about 95% SimpleQA fully local on one RTX 3090 with Qwen3.6-27B
  • Per-user SQLCipher databases; keys derived from the password, never stored
  • No telemetry; Docker images signed with Cosign, with SLSA provenance and SBOMs
  • Downloaded sources build a searchable, embedded personal library

Weaknesses

  • Needs Ollama (or an LLM endpoint) and SearXNG running separately
  • Private or localhost engine URLs are blocked unless an operator env var allows them
  • Requires an AVX-capable x86-64 CPU; older CPUs crash with Illegal instruction
  • docker run --network host only works on native Linux; Docker Desktop needs Compose
  • GPU optional
  • Docker + Compose
  • Needs Ollama or OpenAI-compatible LLM endpoint, SearXNG, SQLCipher (bundled wheels)
  • Models: Ollama models (e.g. gpt-oss:20b, Qwen3.6-27B), any OpenAI-compatible endpoint
  • port 5000
Repository on GitHub ↗ (opens in a new tab)
#964 out of 100

GPT Researcher

Research agent that writes cited reports from web and local documents

30k stars, Apache-2.0, last commit Sep 2026

Planner and execution agents generate research questions, scrape 20+ sources, filter passages (Jev by default, BM25 fallback with no key) and write cited reports over 2,000 words, exportable to PDF and Word. Runs as a FastAPI server on port 8000 with a static or Next.js frontend, or as a pip package; local PDF, Office, CSV and Markdown files can be sources. For analysts automating long-form research.

Strengths

  • Deep Research mode: tree-like exploration, about 5 minutes and $0.40 per run on o3-mini
  • Hybrid retrievers: Tavily plus MCP servers such as GitHub as research sources
  • Works with any OpenAI-compatible endpoint via OPENAI_BASE_URL
  • Multi-agent LangGraph and AG2 variants produce 5-6 page PDF, DOCX and Markdown reports

Weaknesses

  • Default setup needs OpenAI and Tavily API keys
  • Jev context filtering needs a TYPESAFE_API_KEY; the fallback is keyword BM25
  • Python 3.12 or later required
  • Disclaimer labels the project experimental and for academic purposes
  • no GPU
  • Docker + Compose
  • Needs OpenAI or OpenAI-compatible LLM API, Tavily API key (default retriever), TypeSafe API key (optional Jev filter)
  • Models: OpenAI models, any OpenAI-compatible endpoint via OPENAI_BASE_URL, Gemini 2.5 Flash Image (inline images)
  • port 8000
Documentation ↗ (opens in a new tab)Website ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)
#1063 out of 100

PrivateGPT

Anthropic-style API layer for private RAG on local inference servers

57.6k stars, Apache-2.0, last commit Oct 2026

PrivateGPT 1.0 is an API server shaped like the Anthropic Messages API, adding file ingestion, retrieval with citations, web search, code execution, MCP and direct database or CSV querying. It runs no models; it calls any OpenAI-compatible server (Ollama, llama.cpp, vLLM) through OPENAI_API_BASE. A workbench UI at /ui on port 8080 exists for testing; the API is the product.

Strengths

  • Anthropic Messages API shape, so Claude Code, Claude Desktop and Office add-ins can target it
  • Backend-agnostic: any OpenAI-compatible inference server via OPENAI_API_BASE
  • Built-in database and CSV querying, no extra tool server needed
  • Installs with brew or uv tool install; Docker also documented

Weaknesses

  • Runs no models; a separate inference server and embedding server are required
  • No prompt caching and no OAuth or organizations
  • Skills support is marked basic; structured output depends on the backend
  • RBAC, LDAP, connectors and audit logs exist only in the commercial Zylon platform
  • no GPU
  • Docker
  • Needs OpenAI-compatible inference server (Ollama, llama.cpp, vLLM)
  • Models: any model behind an OpenAI-compatible /v1/chat/completions endpoint
  • port 8080
Documentation ↗ (opens in a new tab)Repository on GitHub ↗ (opens in a new tab)