Postgres 16 + pgvector
Semantic search over knowledge runs on an HNSW vector index, inside the same database as everything else.
UNDER THE HOOD
Host it anywhere. Five services. One public origin.
TTSAI ships as one monorepo that builds into five runtime services, and the exact same images bring the whole product up on a managed cloud or on any Docker host. This page is for the person who has to say yes to running it.
5 services, 1 public origin
Postgres 16 + pgvector
Same images, cloud or compose
postgresapiworkermcpweb Only web is public. Everything else is reachable solely over the private network.
ONE ORIGIN
The browser talks to exactly one origin.
nginx serves the app and proxies the API, the WebSocket bridge and the MCP surface on the same origin. Auth cookies stay first-party, and the authenticated app needs no cross-origin grants - the only CORS the API serves is a deliberate, non-credentialed wildcard on the cookie-free public surfaces (widget, forms, portal, public API). TLS terminates at the platform edge, and the site sends HSTS and strict security headers. The embeddable chat widget is the one framing exception, served with a framing-permissive policy so it can live in an iframe on your site.
THE DATA LAYER
Postgres is the platform's spine.
One database carries the records, the vectors, the events and the job queue. Fewer moving parts, one thing to back up.
Force RLS on 101 tables
The app runs as a restricted, non-owner database role, and refuses to boot under any role that could bypass RLS. The database itself is the isolation boundary.
Isolation in depthRealtime, in-Postgres
Writes notify a channel and the API bridges events to WebSockets. The green connection dot in the product header is this bridge.
Background jobs on pg-boss
Inside the same Postgres: scheduled reports, notification digests, nightly pg_dump backups.
Two-phase boot, by design
Migrations run once as the owner role, then the runtime connects restricted. Schema power and runtime power never share a connection.
DEPLOY PATHS
A managed cloud today. Your own metal tomorrow. Same images.
Managed cloud (Railway or similar)
Five services from one repo, private networking, only web exposed, health-checked restarts.
Any Docker host
One compose file brings up the identical images with single-origin parity. Optional profiles add MinIO for attachment storage and Ollama for a fully local LLM. Point the AI at local Ollama and no byte of a customer conversation leaves your infrastructure.
FAILS SOFT
Every external dependency is optional.
No AI key? AI features switch off cleanly and the product says so. No email provider? Outbound mail queues. No object storage? Uploads are disabled with a clear state. No monitoring DSN? A no-op. The core never takes a dependency hostage.
BUILD ON IT
Everything the interface does, the API can do.
REST API, OpenAPI-documented
Scoped Bearer tokens: Read, Write, MCP, Webhook. Secret shown once, individually revocable, last-used tracking.
Signed webhooks
Per brand, with test-fire and a delivery log: status, response code, retries.
Slack & Microsoft Teams
Notifications via incoming webhooks. No code.
MCP for AI agents
Mint a token with the MCP scope and an AI agent can search, create and update work items, governed by the same permissions as any client.
CSV import wizard
Bring existing data in: upload, map columns, run with per-row results.