Nidra
We made Auto Ontology nice, and truly a graph.
Run Nidra
One command on macOS, Linux or Windows, with Docker Desktop.
Run Nidra
One command
macOS and Linux
It asks for your OpenRouter key, starts Nidra and opens it in your browser. Run it again to update; your data stays.
curl -fsSL https://nidra.computer/install.sh | sh
Windows
It asks for your OpenRouter key, starts Nidra and opens it in your browser. Run it again to update; your data stays.
irm https://nidra.computer/install.ps1 | iex
Docker only
Then open the link the last command prints (http://localhost:8086/nidra#key=…). It carries your key, so keep it to yourself.
docker run -d --name nidra --restart unless-stopped \
-p 127.0.0.1:8086:8086 -p 127.0.0.1:3001:3001 -v nidra:/data \
-e OPENROUTER_API_KEY=sk-or-... ghcr.io/sophia-labs/nidra
docker exec nidra nidra url
- Needs Docker Desktop (macOS, Linux or Windows).
- Any OpenAI-compatible key works.
OPENROUTER_API_KEY is the one-variable case. For DeepSeek, pass DEFAULT_MODELS_ENDPOINT=https://api.deepseek.com/v1, DEFAULT_MODELS_API_KEY and DEFAULT_MODELS_MODEL=deepseek-v4-pro.
- The first start catalogues the Chinook sample and compiles its glossary: a few minutes.
- The image is about 730 MB to pull, x86-64; on Apple Silicon, Docker Desktop runs it through emulation.
- It includes the Sophia UI with the Nidra screens and the embedding model. After the pull, only the model calls need the network.
- The image bundles the engine (Apache-2.0), gardend (PolyForm Noncommercial 1.0.0), ONNX Runtime (MIT) and multilingual-e5-small (MIT).
- Prefer Docker Compose and two files? See Advanced.
The graph
What the graph is, for people who do not care about graphs
Auto Ontology started with Neo4j and ported to Postgres; we keep it inside something richer than Neo4j, and even richer than most RDF graphs.
The catalog was always a graph: databases contain schemas, schemas contain tables, tables contain columns, columns point at each other through foreign keys and joins, and a glossary term points at the columns that represent it and the SQL that computes it. What changed is where it is kept.
First: Neo4j
Where Auto Ontology began. Its catalog is a labeled property graph: nodes for databases, schemas, tables, columns, SQL statements and analyses; edges for contains, foreign key, join, union and has-sql.
Then: Postgres
Where NVIDIA moved it. The same graph became Postgres tables, and NVIDIA kept a golden file that captures every read of that graph, to check the port.
Now: gardend
Where Nidra keeps it. The same graph, row for row, as RDF in gardend, the Sophia engine: a direct mapping of the 33 tables, where a row is a resource, a column is a property and a foreign key is a link. NVIDIA’s own golden captures pass identically against it: 1,068 reads. The glossary, the SQL attributes and the vector index are in the same graph; there is no Postgres.
What gardend is
gardend is an RDF graph store: Oxigraph inside, SPARQL 1.1, named graphs. It runs as one process per graph, local-first, and the graph lives in a folder you can copy. On top of the RDF it holds things most RDF stores do not:
find
Search over every entity by meaning and by words, in any language, with the model inside. No API.
Shared documents
Several people and agents can edit the same document at once (a CRDT).
Emporium
A correction is recorded with who said it and when, and its lineage. A disagreement is two heads, not a lost update.
Wires, comments and valuations
On any object. They survive a re-ingest.
Restore points
For documents.
Memory
Agents can recall from a paraphrase.
MCP
Claude or any agent reads and writes the same graph.
What Nidra is
The semantic layer for your warehouse.
Connect a SQL warehouse and Nidra profiles its tables and columns, then compiles a semantic layer over them: a glossary of your business terms, what each column means, the joins that matter and named SQL for your metrics. Every customer gets an isolated knowledge graph that people and agents can read, correct and extend. Ask a question in plain language and you get the answer with the SQL behind it, and your Claude, Cursor or in-house agents can sign in to the same layer over MCP.
The local edition runs on your own machine today, with one command. Hosted Nidra is coming soon. Join the waitlist.
The Nidra engine is open source under the Apache License 2.0; its source is on GitHub: sophia-labs/nidra.
The agent
Ask in plain language, get the SQL
The agent is Auto Ontology’s text-to-SQL agent, kept as it is. It finds the terms and columns a question is about, writes SQL in your warehouse’s dialect, runs it, and repairs the query when it fails. You get the answer, the rows and the SQL behind them.
Its catalog reads go to the graph. Every one of them is graded against upstream’s own recorded answers for the same reads, so the agent sees the catalog it would have seen in Postgres.
What it does
What you get
A catalog of any SQL warehouse
Postgres, Snowflake, Databricks, DuckDB, MySQL, Trino, ClickHouse or SQLite. The ingestion service profiles every table and column and sweeps again every INGEST_INTERVAL_HOURS.
A glossary compiled from it
Business terms, what each column means, the joins that matter between tables, and SQL attributes: named SQL for your metrics, filed under the term they compute.
Answers with their SQL
POST /api/chat/completions streams the agent’s steps, then the answer, the rows and the SQL, so every number can be checked against your warehouse.
One graph for people and agents
The catalog and glossary are RDF in gardend: query them with SPARQL, or attach the engine to your own Sophia app and keep them in a graph of yours.
Hosted Nidra
Hosted Nidra is coming soon
A hosted semantic layer for your warehouse: a login, an isolated graph per customer, an MCP endpoint for your agents, backups and upgrades. Leave your email and we will tell you when it opens. We use this list for that and nothing else.
Early access and questions: vera@sophia-labs.com