Private preview · comparison & deployment reference

Search. Auth.
Realtime. Replicas.
One binary.

KynetraDB brings together a data-plane research catalogue: BM25 full-text search, brute-force vector evaluation, identity, realtime, storage, and replication design. Explore the API, competitor comparisons, operating requirements, and hosting targets in one place.

20Comparison profiles
19Hosting targets
BM25Full-text model
Brute-forceVector evaluation

Each capability has its own page with its state, where it is reached, its limits and the evidence behind it. A capability is marked available when it is reachable and has tests in the engine repository; otherwise its state is lower and says so.

  • 125capabilities
  • 31Available in preview
  • 56Beta
  • 34Roadmap
  • 4Research
  • BM25 full-text search

    Search & retrieval

    Gives an application relevance-ranked keyword search over its own records without running a separate search cluster.

    Available in preview
  • Retrieval evaluation harness

    AI & agents

    Turns retrieval quality into a reproducible number that a build can gate on.

    Available in preview
  • Universal Log

    Data model & realtime

    Gives every write one durable, checksummed home, so a crash or a damaged disk block is detected and handled instead of silently corrupting the database.

    Available in preview
  • Group commit and write durability

    Durability & recovery

    Lets an application choose whether a write is acknowledged after it is on stable storage, while concurrent writers share the cost of each flush.

    Available in preview
  • Sign-up and password sign-in

    Identity & access

    Gives an application accounts and password sign-in without a separate identity service.

    Available in preview
  • Tamper-evident audit log

    Governance & compliance

    Leaves a trail of administrative and governance actions that can be checked for after-the-fact edits.

    Available in preview
  • Discovery document

    APIs, integrations & migration

    Lets tools and other Kynetra products find the interfaces, runtime shape and connection contract of a running instance without hard-coded URLs.

    Available in preview
  • Deployment fit advisor

    Deployment & operations

    Tells you, before you build anything, which KynetraDB deployment mode or interoperability path matches a declared workload, and why other options are ruled out.

    Available in preview

The historical SDK sketch below preserves the API shape that informed KynetraDB’s design. Compatibility is scoped per preview; public parity is not claimed.

quickstart.tstypescript
import { createClient } from '@kynetra/client'

const kdb = createClient('https://api.myapp.com', PUBLISHABLE_KEY)

// Query — fluent builder, fully typed
const { data, error } = await kdb
  .from('posts')
  .select('id, title, author(name)')
  .eq('published', true)
  .order('created_at', { ascending: false })
  .range(0, 19)

// Auth
const { data: { session } } = await kdb.auth.signInWithPassword({ email, password })
const { data: { user } }    = await kdb.auth.getUser()

// Storage
await kdb.storage.from('avatars').upload(`${user.id}.png`, file)
const { data: { publicUrl } } = kdb.storage.from('avatars').getPublicUrl(fileId)

// Realtime — change events + presence + broadcast
kdb.channel('live')
  .on('postgres_changes', { event: 'INSERT', table: 'messages' }, payload => {
    console.log('new message', payload.new)
  })
  .subscribe()

// Edge functions
await kdb.functions.invoke('send-welcome', { body: { userId: user.id } })

// Full-text + vector search
const { data: hits } = await kdb.from('docs').textSearch('body', 'rust async')
const nearest = await kdb.rpc('vector_search', { query: embedding, top_k: 10 })
Reference package
Distribution is scoped during preview

The original client-package shape is retained for architecture review. It is not a public npm installation instruction.

Connect
const kdb = createClient(
  process.env.KYNETRA_URL,
  process.env.KYNETRA_KEY
)

Reference architecture — implementation scope is agreed per preview. The AI feature audit lists one hundred proposed capabilities, each carrying the status a dated audit gave it after naming the source file, API surface and test result behind it. Six of those statuses were re-checked later against engine edition 5f8ea3af and raised; the audit's evidence rows are unchanged. The capability catalogue above is a separate, newer dataset.

The audit changed three of its own entries on review — two down, one up — and those are published as found.See the evidence behind each status → ·Browse all 100 →

Measured, with the artifact attached

  • 2.305 msBM25 top-24 p99 over 5,000,000 documents, against a declared gate of 8 ms.artifact →
  • 8.075 msFiltered vector p99 at 0.1% selectivity — where recall@10 falls to 0.004. Published because tenant-scoped retrieval is exactly this shape.artifact →
  • 7,514 MBResident set for one million entities, about 7,874 bytes each.artifact →

Every figure on this site is local-scope: one host, one operator, no third-party reproduction. Artifacts are linked next to each number, and gates that were missed are published beside the gates that were met.

sdk + rest api

@kynetra/client TypeScript SDK

Fluent typed builder: from().select().eq().order(). Auth, storage, realtime channels, and edge functions — all under one import. REST API at /rest/v1 with select/filter/order/embed/count/upsert/pagination.

replicas

Read replicas + write proxy

Primary writes to a Postgres WAL log; N replicas tail the same log via apply_external, serving reads within ~750ms lag. Write proxy transparently forwards mutations from replica to primary. Lag-based /healthz 503.

target profiles

Deployment target profiles

AWS, GCP, Azure, Fly.io, DigitalOcean, Render, Heroku, Railway, Cloudflare, Oracle, and regional providers are catalogued as operating profiles. Provider automation is scoped during preview.

auth

GoTrue-shaped auth

Email + password, refresh tokens, magic-link/OTP, password reset, OAuth (Google/GitHub), anonymous sign-in. bcrypt (cost 12) + HS256 JWT. Routes at /auth/v1/.

rls

Row-Level Security

auth.uid() policy DSL, evaluated server-side per row for SELECT/INSERT/UPDATE/DELETE. Enforced on the /rest/v1 table routes, KynQL queries, typed views and realtime WebSocket subscriptions; the /v1/collections record routes apply role lists alone, so it is not applied on every data surface. Project service keys skip policies.

realtime

Realtime — WebSocket

postgres_changes (insert/update/delete) + presence + broadcast over WebSocket channels. Only the WebSocket applies row-level policy per subscriber. The server-sent-events stream and the change feed are operator interfaces, not client APIs, and are not filtered per row.

storage

Storage — buckets + signed URLs

Public and private buckets. Signed, time-boxed URLs. Object RLS. Two backends: local disk or any S3-compatible endpoint (AWS S3, R2, MinIO, Wasabi, Spaces). SigV4 built-in. Object keys combine a content-hash prefix with a unique id, so identical uploads are stored separately.

search + vector

BM25 + vector evaluation

BM25 full-text and exact cosine vector search. The preview server scans exactly; the HNSW and SPANN vector indexes are implemented and benchmarked in the engine but are not yet enabled by the server.

operations

Audit log + health probes

Tamper-evident hash-chained audit log (JSONL, /admin/audit/verify). Deep health probes (/healthz, /readyz) with per-check breakdown. Latency ring-buffer → p50/p95/p99 + error rate at /admin/db/metrics.

connect

MCP · CLI · Webhooks · API tokens

MCP server (JSON-RPC 2.0 over stdio) with kynetra_put/get/filter/search/vector_search/stats. Remote CLI with --remote flag. HMAC-SHA256 signed outbound webhooks. SHA-256 hashed API token management.

multitenancy

Organizations → Projects

Organizations → Projects → Members. Each project has its own data namespace, auth pool and API keys. Project data is namespaced by a kind prefix, and project-scoped credentials are confined to their project's namespace. Platform accounts and the global key are operator credentials with deployment-wide authority: issue them only to operators, and keep operator endpoints off the public internet.

agents

Agentic admin

18 registered built-in actions, 3 of them approval-gated, a real LLM tool-use loop (Anthropic + OpenAI + Ollama), persisted run audit trail, and an embedded admin SPA at /admin/v2. No JS build step.

The limits of these cards are stated on their capability pages: row-level security · realtime subscriptions · file storage · vector search · multi-tenancy · built-in agent actions. Tenancy and operator credentials are covered on Security.

Most backend stacks are a fleet of processes — a database, a search engine, a cache, an auth service, a storage API, a realtime broker. Each one is a separate deploy, a separate failure surface, and a separate ops burden.

The original KynetraDB architecture model collapses these concerns into a single data-plane boundary. The diagram is retained as a technical reference:

Your app (Next.js / React / Vue / mobile / CLI / MCP)
      │
      │  @kynetra/client  ·  REST  ·  MCP  ·  CLI
      ▼
kynetradb  ─── one process · one Dockerfile · one WAL
      ├── /rest/v1      query layer (select/filter/order/embed)
      ├── /auth/v1      email + OAuth + JWT + RLS
      ├── /storage/v1   buckets · signed URLs · S3-compatible
      ├── /realtime/v1  WebSocket channels · presence · broadcast
      ├── /functions/v1 WASM edge functions
      ├── BM25 full-text  p50 0.025 ms @ 100k ← built-in
      ├── vector search   exact cosine ← built-in (HNSW + SPANN: engine only, not enabled by the server)
      └── Read replicas   WAL fan-out · staleness p99 17 ms ← built-in

One reference model. One comparison catalogue. 19 target profiles.

Every figure on this site is local-scope: one host, one operator, no third-party reproduction. Artifacts are linked next to each number, and gates that were missed are published beside the gates that were met. See the full benchmark ledger →

Reference catalogue

Architecture scope: historical KynetraDB API, identity, storage, realtime, replication, and operations designs retained for review. PostgREST-style /rest/v1 (select/filter/order/embed/count/upsert/pagination); GoTrue-shaped /auth/v1 (email+password, refresh tokens, magic-link/OTP, OAuth Google/GitHub, anonymous); Row-Level Security with auth.uid() DSL; Realtime WebSocket channels (postgres_changes, presence, broadcast); Storage buckets + signed URLs + object RLS; Edge functions (WASM, sandboxed, 10s timeout); Multitenancy (Orgs → Projects → Members).

Operating model:Read replicas via Postgres WAL fan-out (apply_external, ~750ms lag, /healthz 503 on stale). The deployment catalogue records public-cloud, VPS, Kubernetes, and Indian-provider operating profiles. Provider integrations and automation are scoped during preview.

Search reference:BM25 full-text. Exact cosine vector search (the engine's HNSW and SPANN indexes are not yet enabled by the server). pgvector write-sync + offload opt-in.

Research directions

Enabling indexed vector search (HNSW, SPANN) in the server is not done yet; it is not a shipped public claim.
Auto-rebuild trigger on write threshold — Phase 2.3.
Recall / p99 integration test at 10M vectors — Phase 2.2.
Multi-writer clustering (single primary WAL today) — Phase 3.
Snapshot-and-tail boot (unbounded dataset per node) — Phase 3.
Shard the log (write scale) — Phase 4.
tantivy-grade fuzzy matching + synonyms.
Storage image transforms (width/height/resize).
Provider automation remains preview-scoped and is assessed against the chosen target profile.
Studio policy editor in /admin/v2.

This is a private-preview reference surface. Read the fullState sectionbefore betting anything serious on it.