Neurosymbolic engine
Logic and vectors fused in one: one OWL ontology and one ANN index over the same data.
ABOUT · RESEARCH
The best databases start as research. So did we.
Oxid began as research into neurosymbolic retrieval. Postgres, RocksDB and DuckDB all started the same way: in a lab, as a question worth chasing.
PostgresUC Berkeley
RocksDBresearch
DuckDBCWI Amsterdam
OxidHAN · MADS
THE ORIGIN
Born from research, built for what comes next.
Oxid started life as research. A focused question about whether logical reasoning and vector search could live inside one engine, sharing one identity space, instead of being stitched together by glue code. Creating real reasoning, an artificial brain.
That question turned into a working neurosymbolic retrieval engine. We're proud of where it came from: research is where the hardest ideas get tested before they get shipped. We'd rather earn trust by showing our work than by hiding it.
THESIS
Neurosymbolic retrieval: fusing logical reasoning and vector search in a single engine
"A single identity space lets the ontology filter candidates during the search, not after, eliminating the brittle sync layer between graph and vector stores."
RESEARCH PREVIEW · v0.8.5
A genuinely novel engine.
Oxid today is a strong single-node embedded engine built on a novel neurosymbolic architecture: fast, real, and already useful. It is also early. We'd rather tell you exactly where we are than overclaim and lose your trust.
Novel
Single-node
v0.8.5
ROADMAP
Where we are, and where we're going.
Everything to the right of "Now" is not shipped yet. We label the future as the future, on purpose.
NOWShipped and running today: in the single default binary.
Single-pass hybrid OxQL
Filter by ontology, rank by vector: one query, one pass.
Real reasoning
A live, tractable (PTime) classifier with a commit-time consistency gate.
ACID + WAL durability
Write-ahead log, crash recovery, MVCC snapshots.
Out-of-core storage
Serves datasets far larger than RAM: cold vectors are memory-mapped off-heap, no rebuild needed. Proven serving at ~25× RAM on a single node.
Auto-embedding
Insert text, get vectors: local ONNX, no external service.
Auth, RBAC & TLS
Multi-user accounts, roles, API keys and TLS. Off by default for local dev; opt-in with one command when you expose the server.
Opt-inSingle binary
Embedded graph UI, zero runtime dependencies.
Oxid Surface
The query and tool interface (OxQL and natural language in, reasoned answers out) with a built-in MCP server so agents can call the database as a tool.
COMINGNot shipped yet: actively in progress.
HA · replication · failover
Single-writer, single-node today. Replication and failover to follow.
Biggest gap todayHorizontal scale-out
Out-of-core storage lifts the single-node RAM ceiling; sharding across nodes is next.
Oxid Forge
The agent harness that binds the whole Oxid Ecosystem, remembering, perceiving, speaking and reasoning in a single loop.
VISIONFuture direction: not yet built.
Sensory · sensory-extract
Turn raw documents into structured knowledge automatically at ingest.
Hosted Oxid Cloud
A managed cloud, so you never have to run the engine yourself.
GET STARTED
Read the work. Then come build the future with us.
Dig into the docs to see how the engine works today, or reach out, we love talking to people who value honest, early, well-built systems.