Postgres, Mongo, Redis, Pinecone, Neo4j, Influx, RabbitMQ — replaced by one engine that also answers natural-language questions across your data. For startups that can't afford a 5-person infra team.
red://ask
$ red server --http-bind 127.0.0.1:8080 --path ./data.rdb
red> INSERT INTO hosts (ip, os) VALUES ('10.0.0.1', 'linux');
red> SEARCH SIMILAR TEXT 'suspicious login' COLLECTION logs;
red> ASK 'who owns passport AB1234567 and what services do they use?'
grounded answer
Owner: Alice Costa. Services: billing, admin-console, vpn. Related records were found across table rows, vector matches, graph edges and KV config.
The stack problem
Most startups end up paying five vendors and writing the glue between them. RedDB makes the data model a query capability instead of a separate product to deploy, sync, observe and recover.
fragmented stack
RedDB
ASK
One keyword retrieves rows, documents, vectors, graph edges and KV at once, then calls the model. No retrieval pipeline. No vector store to sync.
ASK 'what vulnerabilities affect host 10.0.0.1?'
RedDB pulls context from SQL rows, graph expansion, document bodies, vector search and KV metadata before generating a grounded response — with provider, sources and prompt token usage attached.
Compare
Five managed services and a custom RAG pipeline, or one engine with native ASK. Same workload, very different surface area.
| Capability | DIY stack Postgres + pgvector + Redis + Pinecone + Neo4j | Supabase Postgres + extensions | RedDB Cloud one engine, 7 models |
|---|---|---|---|
| Relational tables | yes | yes | yes |
| Vector search | pgvector + Pinecone | pgvector | native |
| Graph traversal | Neo4j | no | native |
| Time-series + retention | Influx | no | native |
| Queues / consumer groups | RabbitMQ | no | native |
| Natural-language ASK / RAG | glue code | glue code | one keyword |
| Domain types (IP, GeoPoint, Money, …) | app code | app code | 48 built-in |
| Services to operate | 5+ | 1 | 1 |
| Embedded mode (single binary) | no | no | ~30MB |
| Self-host option | yes | yes | yes (AGPL) |
| Vendor lock-in RedDB Cloud exports backups to your S3. Same file format as self-host. | low | medium | low |
DIY stack
Postgres + pgvector + Redis + Pinecone + Neo4j
Supabase
Postgres + extensions
RedDB Cloud
one engine, 7 models
DIY stack assumes Postgres + pgvector for relational and embeddings, Pinecone for production vector search, Redis for KV / queues, Neo4j for graph traversal, and Influx for time-series. Pricing and ops cost are not in this table — they belong in the TCO calculator.
Supabase row reflects the managed Postgres tier with pgvector and pgmq. Graph traversal, native time-series and queues are listed as “no” because they require app-side modelling or external services.
RedDB Cloud row reflects the engine's public capabilities at v0.2.x. Domain types and ASK are documented on the RedDB docs.
One query surface
RedDB lets collections carry different data semantics without forcing your app to stitch together separate APIs for every model.
Mental model
In RedDB, a collection is a named logical container. A collection can behave like a table, document store, graph, vector index, key-value namespace, time-series series or queue depending on what you write into it.
INSERT INTO users (name, email) VALUES ('Alice', 'alice@co.com')INSERT INTO events DOCUMENT (body) VALUES ('{"level":"warn"}')INSERT INTO identity EDGE (label, from, to) VALUES ('OWNS', 'alice', 'passport:AB1234567')INSERT INTO notes (body) VALUES ('suspicious login') WITH AUTO EMBED (body) USING openaiPUT settings.risk_threshold = 'high'INSERT INTO users (id, name, email) VALUES (42, 'Alice', 'alice@co.com')INSERT INTO logs DOCUMENT (body) VALUES ('{"level":"warn","ip":"10.0.0.1"}')INSERT INTO identity EDGE (label, from, to) VALUES ('OWNS', 'alice', 'passport:AB1234567')INSERT INTO notes (body) VALUES ('suspicious vpn login') WITH AUTO EMBED (body) USING openaiPUT config.risk_threshold = 'high'INSERT INTO cpu_metrics (metric, value, tags) VALUES ('cpu.idle', 95.2, '{"host":"srv1"}')QUEUE PUSH investigations '{"case":"AB1234567","priority":"high"}'SEARCH CONTEXT 'passport:AB1234567' DEPTH 2;
ASK 'who owns passport AB1234567 and what services do they use?';
table · indexed passport match
graph_edge · OWNS passport edge
vector · 0.91 semantic similarity
document · warning login payload
kv · risk_threshold=high
{
"query": "passport:AB1234567",
"tables": [{
"score": 1,
"red_entity_type": "table",
"red_capabilities": ["structured", "table"],
"discovery": { "type": "indexed", "field": "passport" },
"collection": "users",
"entity": { "id": 42, "identity": { "table": "users", "row_id": 42 } }
}],
"graph": {
"nodes": [],
"edges": [{
"score": 0.82,
"red_entity_type": "graph_edge",
"red_capabilities": ["graph", "graph_edge"],
"discovery": { "type": "graph_traversal", "source_id": 42, "edge_type": "adjacent", "depth": 1 },
"collection": "identity",
"entity": { "id": 77, "identity": { "label": "OWNS", "from_node": "alice", "to_node": "passport:AB1234567" } }
}]
},
"vectors": [{
"score": 0.91,
"red_entity_type": "vector",
"red_capabilities": ["embedding", "similarity", "vector"],
"discovery": { "type": "vector_query", "similarity": 0.91 },
"collection": "notes",
"entity": { "id": 91, "data": { "content": "suspicious vpn login" } }
}],
"documents": [{ "collection": "logs", "red_capabilities": ["document", "structured", "table"] }],
"key_values": [{ "collection": "config", "red_entity_type": "kv" }],
"connections": [{
"from_id": 42,
"to_id": 77,
"type": "graph_edge",
"edge_type": "identity",
"weight": 1
}],
"summary": {
"total_entities": 5,
"direct_matches": 3,
"expanded_via_graph": 1,
"expanded_via_cross_refs": 0,
"expanded_via_vector_query": 1,
"collections_searched": 6,
"execution_time_us": 1432,
"tiers_used": ["index", "multimodal", "scan"],
"entities_reindexed": 1
}
}{
"ok": true,
"query": "ASK 'who owns passport AB1234567 and what services do they use?'",
"mode": "sql",
"capability": "table",
"statement": "ask",
"engine": "runtime-ai",
"record_count": 1,
"result": {
"columns": ["answer", "provider", "model", "prompt_tokens", "completion_tokens", "sources_count"],
"records": [{
"values": {
"answer": "Alice Costa owns passport AB1234567. The answer is based on the users collection, the identity graph edge, the logs document, the notes vector match and config key risk_threshold=high.",
"provider": "groq",
"model": "llama-3.3-70b-versatile",
"prompt_tokens": 1834,
"completion_tokens": 74,
"sources_count": 5
},
"nodes": {}, "edges": {}, "paths": [], "vector_results": []
}],
"stats": { "nodes_scanned": 0, "edges_scanned": 0, "rows_scanned": 0, "exec_time_us": 0 }
},
"selection": { "scope": "any" }
}why this matters
The app asks one question. RedDB builds one context set from user rows, evidence documents, graph relationships, semantic matches, configuration state and pending workflow records.
7 models, 1 engine
RedDB keeps one storage format and one query surface while letting each collection behave like the model your workload needs.
01
relational rows and joins
02
JSON records with schema where useful
03
edges, context expansion and traversal
04
semantic search without a sidecar store
05
fast configuration and state
06
retention, downsampling and telemetry
07
FIFO, priority and consumer groups
Types
RedDB ships with 48 built-in types so the app does not parse, normalize and revalidate every IP, coordinate, email, money value, timestamp or cross-model reference.
CREATE TABLE hosts (
ip IpAddr,
cidr Cidr,
owner Email,
location GeoPoint,
monthly_cost Money,
service_ref TableRef
) WITH CONTEXT INDEX ON (ip, owner);
IpAddr · Ipv4 · Ipv6 · MacAddr · Cidr · Subnet · Port
Latitude · Longitude · GeoPoint
Uuid · Email · Url · Phone · Semver
Currency · AssetCode · Money · Decimal
Color · ColorAlpha
Timestamp · TimestampMs · Date · Time · Duration
NodeRef · EdgeRef · VectorRef · RowRef · KeyRef · DocRef
INSERT INTO hosts (ip, cidr, owner, location, monthly_cost, service_ref)
VALUES ('10.0.0.1', '10.0.0.0/24', 'alice@co.com', '-23.550520,-46.633308',
MONEY('USD', 1299, 2), 'services');INSERT INTO hosts (ip, owner) VALUES ('not-an-ip', 'alice');
-- rejected before the record is persisted{
"ok": true,
"query": "SELECT ip, owner, location, monthly_cost, service_ref FROM hosts",
"mode": "sql",
"capability": "table",
"statement": "select",
"engine": "runtime-sql",
"record_count": 1,
"result": {
"columns": ["ip", "owner", "location", "monthly_cost", "service_ref"],
"records": [{
"values": {
"ip": "10.0.0.1",
"owner": "alice@co.com",
"location": "-23.550520,-46.633308",
"monthly_cost": { "asset_code": "USD", "minor_units": 1299, "scale": 2 },
"service_ref": "services"
},
"nodes": {},
"edges": {},
"paths": [],
"vector_results": []
}]
},
"selection": { "scope": "any" }
}No custom parsers for IP, email, CIDR, money, geo and refs in every service.
Validation happens before persistence, so downstream queries do not inherit malformed values.
The engine can route typed values into geo, network, ref and semantic operations.
Clients get predictable JSON shapes for values like Money, vectors and refs.
Embeddings + clustering
RedDB can auto-embed content on write, search semantically, expand context for ASK and cluster vector collections with K-Means or DBSCAN.
INSERT INTO articles (body) VALUES ('AI Safety') WITH AUTO EMBED (body) USING openaiSEARCH SIMILAR TEXT 'suspicious login' COLLECTION logs USING groqSEARCH CONTEXT '192.168.1.1' FIELD ip DEPTH 2Deploy modes
Use it like SQLite inside a Rust app or local tool.
Expose query, admin, backup and operational APIs.
Let AI agents query durable state directly.
AI providers
$ curl -fsSL https://raw.githubusercontent.com/forattini-dev/reddb/main/install.sh | bash
$ npx reddb-cli@latest server --http --bind 127.0.0.1:8080
$ docker run --rm -p 8080:8080 ghcr.io/forattini-dev/reddb:latest
Design partners wanted
We are picking ten startups already running four or more datastores in production and feeling the pain. You get free Cloud Starter for a year. We get the founding-customer feedback that shapes RedDB's first stable release.
What we ask of you
What you get
You can — it works until you also need graphs, time-series with retention, queues with consumer groups, geo, and a RAG pipeline. Then you are wiring five services and writing the glue. RedDB collapses that surface area into one engine and keeps SQL for the workload that wants SQL.
The engine is in production with design partners but the public surface (HTTP, gRPC, file format) can change at minor versions. Use it where moving fast matters more than locking the API. We document every breaking change in the CHANGELOG and ship migration scripts.
read more →AGPL kicks in when you distribute or expose modifications. If you run RedDB as your database behind a SaaS, that is fine — you just cannot fork the engine into a closed-source product. Need a non-AGPL license? Contact sales.
RedDB speaks the Postgres wire protocol, so existing clients (psql, pgx, JDBC, Prisma) connect without changes. Document import is a single CLI command per collection. Vector backfill streams from any OpenAI-compatible embedding API while the new collection is being populated.
Same engine, same file format. Self-host means you run the binary, manage backups, scale yourself. Cloud is the same binary plus a control plane that handles deploys, backups, metrics and updates. There is no Cloud-only feature in the engine.
Self-host: continuous backup hooks for S3, R2, GCS, Turso and D1, plus WAL archiving for point-in-time recovery. Cloud: managed backups, scheduled snapshots and (on the Scale tier) read replicas across regions.
Cloud writes the same .rdb file format as self-host and exports continuous backups to a bucket you own. The CLI imports those backups directly. Worst case: you keep running the open-source build with no migration.
SurrealDB and EdgeDB are multi-model engines but neither ships native ASK / RAG. Supabase is managed Postgres with extensions — great if Postgres is enough for you. RedDB's differentiation is the cross-model ASK plus the seven first-class data models in one engine, with both managed and embedded distribution.
Cache-hit reads land under 1ms. Bulk insert via gRPC sustains roughly 241k ops/sec on commodity hardware. OLAP-heavy GROUP BY work is currently 2–5× behind ClickHouse. Benchmarks and methodology live in the repo.
read more →Self-host: GitHub Issues and the community Discord. Cloud Starter: email with a 24h response target. Cloud Scale: a dedicated Slack/Teams channel and a 99.9% SLA.
Ready to ship one database?
Self-host the open-source build in 60 seconds, or reserve a Cloud Starter slot in private beta.