ollmo

<div align="center"> # ollmo **Open LLM Orchestrator** Your data never leaves — build your own AI productivity tools from your knowledge base. A production-grade RAG platform with hybrid retrieval, multi-tenancy, and streaming chat. Built with Go + Fiber (backend) and Next.js App Router + shadcn/ui (frontend). [Live Demo](https://demo.ollmo.com/) · [Quick Start](#quick-start) · [Documentation](#documentation) · [Roadmap](#roadmap) · [Contributing](#contributing) </div> <div align="center"> English | [简体中文](README.zh-CN.md) </div> ---   ## Why ollmo Most AI application platforms bury you in workflows, nodes, plugins — dozens of concepts and hundreds of settings before the AI can actually help. AI should make things simpler, not the other way around. ollmo converges everything into three steps: **upload documents, ask questions, get answers with sources**. One-command deployment, up and running in ten minutes. - **Simple** — Does one thing and does it well: turn your documents into an assistant you can talk to. - **Easy** — Up in ten minutes, no need to learn a system more complex than your business. - **Restrained** — No concept stacking, no redundant configuration. Complexity stays inside; simplicity stays with you. ## Features - **Hybrid Retrieval** — Dense vectors on Milvus + lexical matching via MySQL FULLTEXT, fused with RRF, with KB-level rerank configuration; built-in retrieval testing with an adjustable keyword ↔ semantic weight slider. - **Multi-Tenancy** — Logical isolation across MySQL, Milvus, and MinIO; every table carries `tenant_id`. - **Document Pipeline** — Upload (drag & drop, paste, multi-file) → parse (MinerU / local parser) → chunk → embed → index, fully async via Asynq + Redis. - **Streaming Chat** — Real-time SSE streaming with source citations. Doubao-style greeting from agent opening message. - **Auto-Memory** — Long conversations are automatically summarized in the background (async LLM task); summaries are injected into later sessions for cross-conversation continuity. Toggleable site-wide. - **Agent Canvas** — Visual agent pipeline (React Flow): intent classification → retrieval → condition branch → LLM → direct reply. Node outputs are referenceable variables (`{node.output}`) in prompts and conditions, enabling LLM chaining and generic branching; per-node debug shows output variables, and nodes highlight live during execution. In-canvas undo/redo and sticky notes. Per-KB system prompt, temperature, top_k, rerank toggle, opening message, and LLM selection. - **Ingestion Pipeline Canvas** — Visual ingestion DAG (React Flow): source → parser → chunker → embedder → sink, with configurable parsing options, chunking strategy, and batch size. - **GraphRAG** — LLM-powered entity & relation extraction at ingestion; entity matching enriches retrieval context at query time. - **Configurable Models** — Tenant-level LLM / Embedding / Rerank provider management via web UI, with per-KB overrides. Selection priority: KB config → tenant default → config.yaml. - **Team & Access Control** — Invitation-based membership with owner / admin / member roles, per-user daily message quotas, KB-level membership (viewer / editor / owner). - **Admin Console** — Super-admin user management, tenant management with plan & quota control (free / pro / enterprise), system settings (registration toggle, auto-memory), analytics dashboard, and audit logging. - **External API** — API-key authenticated programmatic access (search, KB management, streaming chat). - **Internationalization** — Chinese / English UI with server-side message contracts. No locale in URL paths. ## Tech Stack | Layer | Choice | |---|---| | Backend | Go 1.26+ / Fiber / GORM | | Frontend | Next.js 15 (App Router) / React / shadcn-ui | | Database | MySQL 8.0 (utf8mb4) — metadata | | Vector DB | Milvus 2.5+ — dense vector retrieval | | Object Storage | MinIO (S3 compatible) — original docs, parsed output, thumbnails | | Task Queue | Asynq + Redis — parsing, embedding, index rebuild, auto-summarization | | Document Parser | MinerU (containerized, GPU optional) + local fallback parser | | Auth | Local account (email + password + JWT) | ## Architecture ``` ┌─────────────────────────────────────────────────────┐ │ Next.js Web (3001) │ └────────────────────────┬────────────────────────────┘ │ HTTP / SSE ┌────────────────────────▼────────────────────────────┐ │ Go API Server (8080) │ │ Fiber · JWT auth · Tenant isolation · RESTful API │ └───────┬────────────┬──────────────┬─────────────────┘ │ │ │ ┌────────▼──┐ ┌──────▼──────┐ ┌───▼────────────┐ │ MySQL │ │ Redis │ │ MinIO │ │ metadata │ │ task queue │ │ file storage │ └───────────┘ └──────┬──────┘ └────────────────┘ │ ┌────────▼────────┐ │ Asynq Worker │ │ parse · embed · │ │ summarize │ └────────┬────────┘ │ ┌────────▼────────┐ │ Milvus │ │ dense (vectors) │ └─────────────────┘ ``` ## Project Layout ``` ollmo/ ├── server/ # Go backend (API + worker share the same binary) │ ├── main.go # Entry point: `go run . api|worker|migrate` │ ├── internal/ # Business domains (one package per domain) │ │ ├── agent/ # Agent canvas config & execution │ │ ├── analytics/ # Usage statistics & activity feed │ │ ├── apikey/ # API key management & middleware │ │ ├── audit/ # Audit logging │ │ ├── auth/ # Authentication (email + password + JWT) │ │ ├── chat/ # Streaming chat with citations │ │ ├── doc/ # Document lifecycle, chunking, worker │ │ ├── embedding/ # Embedding model management │ │ ├── graph/ # GraphRAG (entity extraction & retrieval) │ │ ├── invitation/ # Team invitation system │ │ ├── kb/ # Knowledge base CRUD │ │ ├── llm/ # LLM model management │ │ ├── memory/ # Auto-summarization & cross-session context │ │ ├── pipeline/ # Ingestion pipeline canvas config │ │ ├── rerank/ # Rerank model management │ │ ├── search/ # Hybrid retrieval (dense + lexical + rerank) │ │ ├── site/ # Site-wide settings │ │ ├── team/ # KB membership & access control │ │ ├── tenant/ # Tenant, plan & quota management │ │ ├── user/ # User management │ │ └── ... │ ├── pkg/ # Shared infrastructure (db, vector, clients) │ └── config.yaml ├── web/ # Next.js frontend │ ├── app/ # App Router pages ('/' chat, '/dashboard/*' admin) │ ├── components/ # UI + feature components (agent, pipeline, etc.) │ ├── lib/ # API client, i18n, utils │ └── messages/ # i18n message files (zh, en) ├── deploy/ # Dockerfiles ├── docker-compose.yml # Full-stack local deployment └── Makefile # Dev & build targets ``` ## Quick Start ### Prerequisites - [Docker](https://docs.docker.com/get-docker/) + Docker Compose - [Go](https://go.dev/dl/) 1.26+ - [Node.js](https://nodejs.org/) 20+ - [Make](https://www.gnu.org/software/make/) (optional, for convenience targets) ### 1. Clone & configure ```bash git clone https://github.com/ollmo-go/ollmo.git cd ollmo cp .env.example .env ``` ### 2. Start infrastructure ```bash make up ``` This starts MySQL, Redis, MinIO, Milvus, and the Asynqmon dashboard. | Service | URL | Credentials | |---------|-----|-------------| | MySQL | `localhost:3306` | `ollmo` / `ollmo_dev_pwd` | | MinIO Console | http://localhost:9001 | `minioadmin` / `minioadmin` | | Milvus | `localhost:19530` | — | | Asynqmon UI | http://localhost:8081 | — | ### 3. Run database migrations ```bash make migrate ``` ### 4. Start backend (API + worker) ```bash # Terminal 1: API server make server # Terminal 2: Asynq worker (parsing, embedding, auto-summarization) make worker ``` ### 5. Start frontend ```bash make web # http://localhost:3001 ``` ### Full stack via Docker ```bash make up-full # Builds & starts everything (infra + server + worker + web) ``` ## Configuration | Source | Purpose | |--------|---------| | `server/config.yaml` | Backend defaults (DB, Redis, MinIO, Milvus, MinerU, JWT, quotas) | | `.env` | Environment overrides for Docker deployment | | Web UI → Settings | Per-tenant LLM / Embedding / Rerank models (with default selection) | | Web UI → KB Settings | Per-KB embedding / rerank model override | | Web UI → Agent Canvas | Per-KB agent config (system prompt, LLM, retrieval params, opening message) | | Web UI → Pipeline Canvas | Per-KB ingestion config (parser, chunker, batch size) | | Web UI → System Settings | Site-wide settings (registration toggle, auto-memory) — super admin only | Model selection priority: KB-level config → tenant default → config.yaml fallback. ## Development - All development happens on the `dev` branch. `main` is release-only. - Modules are organized by **business domain**, not technical layer. Each domain package contains its own `model.go` / `repo.go` / `service.go` / `handler.go`. - Comments are kept in English and minimal — only where the intent is non-obvious. ```bash make build # Build server binary + web make build-docker # Build Docker images make clean # Remove build artifacts ``` ## Documentation - [Live Demo](https://demo.ollmo.com/) — Try it before you deploy. ## Roadmap - [x] **Phase 1 — MVP**: KB/document CRUD, document parsing, embedding, hybrid retrieval, streaming chat with citation, local auth + tenant isolation. - [x] **Phase 2 — Configurability**: Multi LLM/Embedding/Rerank providers, KB editing with re-indexing, rerank, multi chunking strategies, drag & drop upload, i18n. - [x] **Phase 3 — Intelligence**: Ingestion Pipeline canvas (React Flow), GraphRAG, Agent canvas, Memory, role-based access control, external API, team management. - [x] **Phase 4 — Production Readiness**: Analytics dashboard, audit logging, conversation management (search / rename / export / pin), auto-memory summarization, super-admin console (users, tenants, plans & quotas, system settings), registration toggle. - [ ] **Phase 5 — Document Intelligence**: ### 5.1 Document Preview & Citation Tracing - In-chat citation clicks jump to the source document with the cited passage highlighted - Document viewer: PDF render, markdown/text preview, page navigation - Chunk-to-source mapping (chunk → document → page/offset) - Side-by-side view: chat + source document ### 5.2 Advanced Document Parsing - Table extraction: preserve table structure as HTML for structured retrieval - Image extraction: store images in MinIO, embed captions, support vision LLM Q&A - Layout-aware parsing: recognize headers, footers, columns, footnotes - OCR improvement: scanned document support with language detection ### 5.3 External API Expansion - Full CRUD via API key: upload documents, create/manage KBs, list conversations - OpenAPI/Swagger documentation at `/api/v1/docs` - Rate limiting per API key (requests/min, tokens/day) - Webhook callbacks for document processing events (parsed, embedded, failed) ## Contributing 1. Fork the repository and create your branch from `dev`. 2. Ensure `make build` passes before submitting a PR. 3. Keep changes focused — one feature or fix per PR. 4. Follow existing code style: English comments, domain-driven package structure. ## License Apache License 2.0 with additional conditions. See [LICENSE](LICENSE) for details. In short: - ✅ Free for personal use, internal enterprise use, and commercial application development. - ❌ Operating a multi-tenant SaaS service similar to ollmo requires a commercial license. - ❌ Removing or modifying the ollmo brand, LOGO, or copyright notices in the console is prohibited.