One endpoint. Five query patterns. One engine. Rows, vector
embeddings, full-text postings and graph edges commit together in one
atomic write - every write instantly visible to SQL, vector, graph,
full-text and the natural-language /ask alike.
SELECT symbol, close FROM prices WHERE date > $1 01 topk(embedding, :q, 10) WHERE sector = 'tech' 02 MATCH (a)-[:REFERRED]->(b) RETURN b 03 search('quarterly earnings beat') 04 /ask "top movers this week" 05 
A member of the NVIDIA Inception program and the AI Council of India (IAMAI). Learn more →
Most modern AI stacks fan a single user action across four databases - a relational store, a vector index, a search engine, and a graph. Each has its own bearer, its own SDK, its own billing line, and its own way of being out of sync at 3 am. OriginChain replaces all four with one managed database. One write lands every shape atomically.
Retire four contracts. Onboard one. Stop coordinating four maintenance windows for a single feature release.
The row, its embedding, its full-text postings, and its graph edges commit together in one atomic write. No reconciliation jobs. No "the search index is 8 hours behind" disclaimers.
One SDK, one auth model, one observability surface. New AI features land without a Debezium pipeline + four retry queues + a 3 am reconciliation cron.
SQL · Vector · Full-text · Graph · Natural language, all on the same store and the same bearer. Write your data once, query it four ways without bolting on a separate engine.
p99 under 8 ms for typed SQL. HNSW vector top-k at 100k vectors: recall@10 = 0.96 with p99 109 ms in default high_recall mode, or p99 37 ms in fast mode (recall 0.69). At scale, IVF-PQ reaches recall@10 0.979 on the 100M-vector BIGANN benchmark at p99 333 ms on a single box. Warm natural-language queries return in under 50 ms thanks to plan caching.
Your own compute, your own volume, your own bearer - locked to the region you pick. Zero shared resources, zero noisy neighbours, zero cross-tenant blast radius.
Provisioning, TLS, backups, replication, and observability run for you. You ship product features; we run the database underneath them.
POST a sentence to /ask and get rows back. Plan-cached natural-language queries execute in-region with sub-50 ms warm latency and a stable schema-aware response shape.
RPO=0 on higher configurations and ~25 second failover keep your application online through hardware loss. Writes are continuously archived to encrypted object storage, so you can restore to any timestamp.
Concrete p99 numbers, not averages - the ceiling your app can plan around for SLAs, agent loops, and user-facing reads on a managed database.
End-to-end p99 in-region for typed queries against your managed database.
After the first ask of a shape, the plan is durably cached and every repeat skips compile.
recall@10 = 0.979 at 100M on BIGANN (real data, published ground truth) — DiskANN-class leader-band accuracy, on a single box.
First ask of a brand-new shape from another continent - including the round trip.
p99 measured at the API edge · in-region unless noted · vector topk is the 100M-vector benchmark — IVF-PQ on BIGANN (real data, published ground truth), recall@10 = 0.979 at p99 333 ms on a single box · the HNSW path serves smaller corpora at single-digit-millisecond p99
OriginChain is the AI-native database for teams that need SQL, vector, full-text, graph, and natural-language queries on one managed endpoint. Single tenant, region-isolated, fully hosted - provision in under two minutes, query in milliseconds.
For decades a database had one job: exact answers to precise queries. The AI era needs more — semantic flexibility and deterministic precision, on one consistent copy of your data. Why we built OriginChain as a true multi-modal database.
The modern AI stack is a database plus a vector DB plus a search index plus glue code. OriginChain is one managed store that answers SQL, vector, full-text, graph, and natural-language queries — the same write visible to every shape, atomically.
We refactored our multi-region DR from a textbook hot-standby model to a cold-standby model and dropped per-tenant DR cost from ~$78/mo to ~$5–15/mo. RPO ≤60 s, RTO ~10–15 min. The math, the runbook excerpt, and the trade-off named explicitly.
Pick a region, choose your configuration, and we provision a dedicated single-tenant instance in roughly ninety seconds. The first query you send is the first query we'll show you how to write - in English.