// the find
FalkorDB/QueryWeaver
An open-source Text2SQL tool that transforms natural language into SQL using graph-powered schema understanding. Ask your database questions in plain English, QueryWeaver handles the weaving.
QueryWeaver is a Python service that turns plain-English questions into SQL for PostgreSQL and MySQL databases. It keeps each schema as a graph in FalkorDB, so the model works from tables, columns and foreign keys rather than raw DDL. It ships as a Docker image with a REST API, an MCP endpoint and a Python SDK, which suits teams that want a self-hosted query assistant rather than a hosted BI tool.
The schema is stored as a graph, and a separate relevancy agent (api/agents/relevancy_agent.py) picks what matters before SQL is written. That split is the right one for large schemas, where pasting all the DDL into the prompt stops working. Destructive statements are gated: INSERT, UPDATE and DELETE stop at a confirmation step before anything runs, and tests/test_destructive_detection.py and tests/test_sql_sanitizer.py cover that path. The auth design is written down in unusual detail. Browser sessions are signed cookies that survive a FalkorDB outage, API tokens are checked server-side and can be revoked one at a time, and email signup creates no account until the six-digit code is typed back in the same browser. The Python SDK accepts per-request LLM overrides and keeps no global state, so several instances can run side by side in one process.
FalkorDB is a hard dependency even at boot. The README says the process will not start without it, so every database it serves sits behind one more service. An Anthropic-only setup also fails at startup unless you add a Voyage key or set EMBEDDING_MODEL, because Anthropic has no embeddings endpoint. That is the kind of thing that bites on the first run. The license is AGPL, so anyone who runs a modified copy as a network service has to publish their changes, which will rule it out for some companies that want to embed it in a product. The README has no accuracy numbers or benchmark on messy real-world schemas, so how often the generated SQL is right is something you would have to measure yourself before trusting it.