Service status
Measured over the last 7 days — synthetic probes every 5 minutes plus every real request. Machine-readable at /api/status.
Service-level objectives
| Availability (synthetic probes) | target 99.9% | measured 100% | met |
| Search latency p95 (all real search calls) | target < 300 ms | measured — | measuring |
| Error rate (all surfaces) | target < 1% | measured 99.68% | missed |
2016 probes since 2026-09-28 21:30 UTC · DB p50 32 ms · search-probe p95 122 ms
Per-surface traffic (7d)
| Surface | Calls | Errors | p50 | p95 | p99 |
|---|---|---|---|---|---|
| MCP (agents) | 0 | 0% | 0 ms | 0 ms | 0 ms |
| REST API | 0 | 0% | 0 ms | 0 ms | 0 ms |
| oasy landing (crawlers) | 57,813 | 99.68% | 144 ms | 4473 ms | 19933 ms |
How this platform scales
- Autoscaling serverless compute (Vercel), pinned to
cdg1(Paris) — colocated with the database (AWS eu-west-3) so every query is a single-digit-millisecond hop. - Managed Postgres (Supabase) behind pgbouncer connection pooling; transaction pooler for request paths, session pooler for migrations.
- Retrieval infrastructure operated for the publisher: chunking + embeddings written at ingest, HNSW vector index, hybrid semantic/full-text search with a latency budget and a two-tier query-embedding cache.
- Model compatibility: every stored vector records its embedding model; search only compares vectors from the active model, so model upgrades re-embed incrementally with zero downtime — and content is consumable by any client over MCP, A2A, or plain REST.
- Edge & token gateway: per-publisher rate limits, OAuth 2.1 bearer tokens hashed at rest, WAF/CDN redirect patterns for crawler traffic, custom-domain routing.