README.md
</picture>
<h1 align="center">Wren AI - Open-Source GenBI Agent</h1>
<p align="center"> </p>Ask your database anything in plain English. Wren AI generates accurate SQL, charts, and BI insights β backed by a semantic layer that keeps LLM outputs grounded and trustworthy.
https://github.com/user-attachments/assets/f9c1cb34-5a95-4580-8890-ec9644da4160
βΆοΈ Watch the full GenBI walkthrough β end-to-end from question to chart
Feeding raw DDL to an LLM gets you SQL that looks right but means the wrong thing β "revenue" joins the wrong tables, "active user" uses the wrong filter. Wren AI's semantic layer (MDL) encodes your business definitions once, then every generated query is grounded in that shared understanding. The LLM doesn't guess what your metrics mean; the semantic layer tells it.
| What you get | Why it matters | |
|---|---|---|
| Talk to Your Data | Ask in any language β precise SQL & answers | Slash the SQL learning curveο»Ώ |
| GenBI Insights | AI-written summaries, charts & reports | Decision-ready context in one clickο»Ώ |
| Semantic Layer | MDL models encode schema, metrics, joins | Keeps LLM outputs accurate & governedο»Ώ |
| Embed via API | Generate queries & charts inside your apps (API Docs) | Build custom agents, SaaS features, chatbotsο»Ώ (Streamlit Live Demo) |
| Cloud Warehouses | Databases | Query Engines |
|---|---|---|
| BigQuery | PostgreSQL | Trino |
| Snowflake | MySQL | Athena (Trino) |
| Redshift | Microsoft SQL Server | DuckDB |
| Databricks | ClickHouse | |
| Oracle |
Don't see yours? Vote for it β community votes drive our connector roadmap.
Wren AI works with any LLM provider you're already using:
| Cloud APIs | Platform Services | Self-hosted |
|---|---|---|
| OpenAI | Azure OpenAI | Ollama |
| Anthropic | Google AI Studio (Gemini) | |
| DeepSeek | Vertex AI (Gemini + Anthropic) | |
| Groq | AWS Bedrock | |
| Databricks |
[!TIP] For best results, use a frontier model (GPT-4o, Claude Sonnet, Gemini Pro). Wren AI works with smaller and local models too β accuracy scales with model capability. See configuration examples for setup guides.
Three ways to get started β pick what fits:
| Option | Best for | Link |
|---|---|---|
| Self-hosted (Docker) | Full control, local data | Installation guide |
| Wren AI Cloud | Try it without setup | getwren.ai |
Compare OSS vs. Cloud plans. Full documentation at docs.getwren.ai.
<p align="center"> </p>User questions flow from the Next.js UI β Apollo GraphQL β AI Service (RAG + LLM) β Wren Engine (semantic query execution) β your database. The semantic layer (MDL) sits at the center, making sure the LLM's SQL reflects your actual business definitions.
π Deep dive into the design
WrenAI is a full-stack AI system with interesting problems at every layer β semantic modeling, RAG retrieval, LLM-driven SQL generation, and query execution across heterogeneous data sources. Here's what the stack actually looks like under the hood:
| Layer | What it does |
|---|---|
| wren-ui | Next.js + Apollo GraphQL β semantic modeling UI and the BFF that wires everything together |
| wren-ai-service | Python/FastAPI pipeline β intent classification, vector retrieval from Qdrant, LLM prompting, and SQL correction loops |
| wren-engine | Rust + Apache DataFusion β the query execution core that resolves MDL semantics (metrics, joins, access controls) before SQL reaches the database |
wren-engine is a separate open-source project and the part of the stack closest to the metal. It's where MDL definitions get translated into actual query plans across 15+ data sources. If you work with Rust, DataFusion, or database connectors, it's worth a look β the codebase is approachable and there are real unsolved problems around query planning, semantic resolution, and MCP (Model Context Protocol) agent integration.
Some areas where contributions tend to have the most impact across both repos:
We follow a Code of Conduct to keep the community welcoming for everyone.