Frequently asked questions

OriginChain is a managed AI-native database. You write a question in English, the instance compiles it into a query plan once, caches it, and serves every repeat in milliseconds. SQL, vector search, full-text, and graph traversal all run against the same store, behind a single bearer and a single endpoint.

A single-tenant OriginChain instance in the region you pick: an HTTPS endpoint, a bearer token, a private write-ahead log, and encrypted nightly backups. We provision, operate, TLS-terminate, rate-limit, back up, and upgrade. You post requests, you get JSON.

Tier 1 starts at $210/mo for first production workloads, Tier 2 is $699/mo with a 99.9% SLA, and Tier 3 is $2899/mo with a 99.95% SLA for high-throughput deployments. Storage is billed separately, $9 to $699/mo depending on capacity. Add-ons (SQL Pro, Vector Search, Full-Text Pro, Graph, Transactions, sub-second PITR) attach to any tier. Annual billing saves 17%. Enterprise is custom - contact sales.

No free trial. Every account adds a card at signup and is billed from day 1 on whichever tier you pick. Tier 1 starts at $210/mo and gives you a real production endpoint with real backups - no sandbox. Cancel any time and we prorate the unused days. Enterprise pilots are structured per-account; talk to sales.

Asia Pacific (Mumbai) is live today. US East, US West, EU Ireland, EU Frankfurt, AP Singapore, AP Tokyo, and South America (São Paulo) are coming soon - each one lights up as AI compute is brought online there. Your data, write-ahead log, backups, and the AI compile call all stay inside the region you pick.

Sign up at app.originchain.ai. Pick a region, pick a compute and storage tier, enter a card. In about two minutes you receive an HTTPS endpoint and a bearer token. From there: install the SDK or call the HTTP API directly, declare a collection, and post your first row. Quickstart docs are at originchain.ai/docs.

The marketing site (originchain.ai) and the dashboard (app.originchain.ai) are separate apps with separate deploy cadences. The marketing site has no auth surface, no cookies, and is aggressively SEO + CDN cacheable. The dashboard is an authenticated SPA with proper CSP and per-tenant routing. This is the same split Stripe, Linear, and Neon use - it keeps auth surface, cookie scope, and analytics cleanly separated.

Six surfaces against one store: SQL (correlated EXISTS, aggregate over expression, CASE WHEN, UNION/INTERSECT/EXCEPT, predicate pushdown to secondary indexes, EXPLAIN ANALYZE with per-node cost), vector search (HNSW with cosine, inner product, L2, and Manhattan; sparse vectors; PQ scaffold; adaptive over-fetch with predicate filters), full-text (BM25, phrase queries, UAX #29 Unicode tokenizer, 18-language stemming), graph (9 algorithms - multi-hop traversal, filtered traversal, variable-length paths, shortest-path BFS, connected components, triangle enumeration, all-simple-paths forward + bidirectional, PageRank), natural-language /ask, and reactive /watch over the WAL. Every shape compiles down to reads against the same managed engine.

Every write across every shape (SQL row, vector embedding, full-text postings, graph edge) commits in one WAL frame on one substrate. Either every shape sees the new row, or none of them do - by construction, not by best-effort. There's no dual-write coordinator, no eventual-consistency window between a row and its embedding, no ETL job to sync your vector DB with Postgres. The substrate is a single hash-keyed k/v store; every shape is a key shape on top.

Yes. A row's embedding lives at a vector key prefix in the same store as the row itself. Atomic batched writes mean your row, its embedding, and its full-text postings always commit together - no separate vector DB to ETL into, no eventual-consistency window between systems.

Send an English sentence to /ask and the instance compiles it into a query plan once, caches the plan, and reuses it for every future request that matches that shape. The compile step calls the LLM you configure (via BYOK LLM) only on first touch. After that, the LLM is not consulted again - repeat asks return from the plan cache in milliseconds. Cost-walker-informed planning uses real cardinality from oc-stats so the LLM picks from a shortlist that already passes a cost test.

Yes. Configure an OpenAI, Anthropic, Gemini, or Groq key per tenant; /ask routes through it. We never bill you for LLM tokens - your contract, your audit, your bill. Keys are envelope-encrypted at rest and isolated per tenant. Every call is audited (provider, model, prompt tokens, completion tokens) and visible in /usage. A fallback provider can be pinned for transparent retry on error or timeout.

Only the English sentence and a small schema-shape catalog snippet (column names and types). It never sees rows, never sees the WAL, never sees secrets. The compile call stays inside the same region as your instance.

Yes - nine algorithms shipped. Multi-hop traversal, filtered traversal with per-hop WHERE predicates, variable-length paths (*1..N), shortest-path BFS, connected components via Union-Find, triangle enumeration, all-simple-paths forward and bidirectional (5-20x faster on real graphs), and PageRank via power-method iteration. Per-hop EXPLAIN ANALYZE shows frontier size, edges considered, and time per hop in a single response. All on the same store as your rows - no separate graph DB.

The default tokenizer is UAX #29 Unicode-segmented - it handles Latin, Cyrillic, Greek, CJK (Chinese, Japanese, Korean), Arabic, Devanagari, and Hebrew out of the box. The optional analyzer pipeline adds Snowball stemming for 18 languages (English, French, German, Spanish, Italian, Portuguese, Dutch, Russian, Swedish, Norwegian, Danish, Finnish, Hungarian, Romanian, Turkish, Arabic, Tamil, Indonesian) plus diacritics fold and stop-words. BM25 ranking uses Lucene defaults (k1=1.2, b=0.75).

Single-row writes are atomic and single-row CAS gives you optimistic concurrency control via _oc_row_version. That covers most OLTP patterns. For multi-row snapshot-isolation transactions, attach the Transactions add-on - it gives you begin / get / put / delete / commit / abort across multiple rows under snapshot isolation.

Cached query path is sub-millisecond inside the engine plus a few milliseconds over the network - Tier 2 and Tier 3 hit p99 under 8 ms in-region. First-time /ask queries round-trip to the AI compiler and typically return in 200–500 ms; that cost is paid once per question shape and every subsequent variant of the same shape is on the fast path.

Cached intents return in under 50 ms. First-time queries compile via the LLM you configure (BYOK LLM) and typically return in 200-500 ms depending on schema size and sentence complexity. Schema prompt-caching gives you a ~90% discount on repeated catalog tokens.

Four metrics x four index variants: cosine, inner product, L2, and Manhattan (L1) across dense HNSW, sparse vectors, Product Quantization (scaffold), and adaptive over-fetch with predicate filters. HNSW topk ships two modes - default `high_recall` hits recall@10 = 0.96 at 100k vectors with p99 109 ms; `fast` mode runs p99 37 ms at recall 0.69 (use it for RAG with re-rankers or hot dashboards). Filtered topk uses metadata equality predicates evaluated during graph traversal. Brute-force fallback for very small indexes. Numbers above are from a deterministic 100k-vector benchmark (D=128, M=16, ef_construction=200, 1000 queries).

Tier 1 targets ~100 req/s sustained and 50M rows. Tier 2 targets ~1,000 req/s sustained and 500M rows with a 99.9% SLA. Tier 3 targets ~5,000 req/s sustained and 2000M rows with a 99.95% SLA. You resize compute and storage independently - both hot, no downtime.

HNSW is verified at 100k vectors per index with the recall and latency numbers above (recall@10 = 0.96, p99 109 ms in default high_recall mode). Many tenants run several indexes per instance for hundreds of thousands of total vectors. For larger single-index workloads, contact sales@originchain.ai about Enterprise - we'll match the index strategy to your dataset.

Every managed instance ships with encrypted nightly backups at no extra cost. Backups stay inside the region of your instance and are integrity-checked daily. Restore is available on demand from the console or the HTTP API.

Yes. Continuous-archive PITR is included on every paid tier - restore to any 15-minute checkpoint from the dashboard or the HTTP API. Sub-second point-in-time recovery is available as an add-on for compliance-heavy workloads that need to rewind to a specific event.

Active-passive with synchronous replication. RPO = 0 on paid tiers - every commit is acknowledged by the standby before the client sees success. RTO is around 25 seconds in our drills. Snapshot-bootstrap means a new follower can join without operator intervention, and failover preserves full state.

Per-instance metrics (request rate, latency percentiles, error rate, plan-cache hit rate, replication lag) in the console, plus OTLP push to your own collector if you bring an endpoint. Audit logging is on by default and queryable through the console or the API. EXPLAIN is available on every query path.

Python is available today - fully typed, both sync and async, with helpers for /rows, /query, /ask, /watch, /sql, /vector, /fts, and /graph. Coming soon: TypeScript. The HTTP surface is documented and stable, so any language with an HTTP client works today.

Yes. The /v1/watch endpoint streams updates over Server-Sent Events - WAL-native, so every write puts an event on the stream by construction. No separate CDC pipeline, no Debezium, no replication lag. Prefix subscriptions ship today; predicate-filtered watches land this sprint. Per-tenant concurrent-subscriptions cap with RAII guard - client disconnect releases the slot the same tick.

Yes. Every instance runs on a dedicated, region-isolated host. No shared compute, no shared storage, no noisy neighbors. Your data, your write-ahead log, your backups, and the AI compile call all stay inside the region you pick.

TLS 1.3 in transit on every endpoint. AES-256 encryption at rest for the data volume and for backups. Bearer tokens authenticate every request and rotate via the console with a 60-second grace window for rolling deploys. Customer-managed keys (BYOK) are available on Enterprise.

Yes. Every authenticated request is recorded with timestamp, principal, endpoint, and outcome. Audit logs are queryable through the console and the API, exportable to your own SIEM via OTLP, and retained according to your tier's retention policy.

SOC 2 Type 1 audit is underway with our auditor - contact security@originchain.ai for the current letter and timeline. HIPAA BAA and GDPR DPA are available on Enterprise. Every instance ships with TLS 1.3, bearer auth, encrypted backups, single-tenant isolation, and audit logging regardless of tier.

Inside the region you pick - and only that region. Compute, storage, write-ahead log, backups, audit logs, and the AI compile call all stay in-region. You can pin Enterprise instances to a specific availability zone for residency contracts that require it.

Add-ons are flat-monthly capabilities you attach to any instance - SQL Pro ($49), Vector Search ($79), Full-Text Pro ($49), Graph ($59), Transactions ($99), Intra-Segment PITR ($149), Multi-Writer Cluster (Enterprise). Toggle them on or off any time from the billing page; charges are prorated by the day.

Upgrades and add-on attachments are prorated by the day on your next invoice. Downgrades and detachments take effect immediately and credit the unused portion of the cycle. Resizing compute or storage is a hot operation - no downtime, no reconnection needed.

Yes. Annual billing saves 15% on compute, storage, and add-ons. Talk to sales@originchain.ai for multi-year terms or volume commitments.

Overage is charged on your next cycle at $0.02 per natural-language query, $0.0001 per DB call, and $0.12 per GB of egress above the tier baseline. You see a real-time meter in the console, and you can set hard caps that throttle requests with a 429 once a threshold is hit.

Custom SLA, customer-managed keys (BYOK), HIPAA BAA, GDPR DPA, multi-region deployments, named support engineer, negotiated quotas, custom retention, multi-writer cluster access, and an annual contract. Contact sales@originchain.ai.