System requirements
PoC profile
| Item | Reference profile |
|---|---|
| Machine | One x64 VM, 4 vCPU, 16 GB RAM |
| Disk | 100 GB SSD |
| System | Linux x86_64 (Ubuntu 22.04 LTS, Debian 12 or later) or Windows Server 2022 x64 or later: minimums |
| Database | PostgreSQL 18, pgvector built for PG18, unaccent extension |
| Files | Managed SeaweedFS, binary included in the Backend package |
| Models | Remote conversation, embedding, OCR and image services |
| User access | VM DNS name, HTTPS to the Frontend service on TCP 9000 |
| Administration | sudo on Linux or an elevated Windows console for installation |
This profile defines the documented scenario; it does not guarantee a simultaneous user count. No GPU is required on the VM because inference is remote. The PoC journey gives the installation sequence.
Operating systems
| System | Minimum required | Validated |
|---|---|---|
| Linux | x86_64, glibc 2.35 or later, systemd: Ubuntu 22.04 LTS, Debian 12 or later | Ubuntu 24.04 LTS |
| Windows | Windows Server 2022 x64 or later, 64-bit PowerShell 5.1 | Windows Server 2022 |
Linux binaries are built against glibc 2.35: RHEL 9 and its derivatives (glibc 2.34) are not supported.
Host dependencies
| Linux | Windows |
|---|---|
| systemd | Windows services and Windows Firewall; NSSM as an alternative mode |
curl, for the procedure checks |
Visual Studio C++ tools, to build pgvector on the server or on a preparation workstation |
| Prepare PostgreSQL | Prepare PostgreSQL |
SmartAGT executables bundle their runtime: no Python needs to be installed on the server. On Linux, the Backend package also bundles libmagic in its magic/ folder and uses it even if the host has its own. To use the system one, install its package and remove libmagic.so.1 and magic.mgc from this folder.
Network and certificates
Private network
SmartAGT is designed for a private network, without direct Internet exposure: Network exposure.
On a server, the Frontend service, the Public API and the Gateway respond by default over HTTPS with a self-signed certificate generated by SmartAGT. In a cluster, every Frontend node receives its certificate during installation. For lasting use, prepare your own certificates, trusted by workstations and applications: Install your certificates.
Also prepare trust for connected external services. Encrypting PostgreSQL is recommended in production: PostgreSQL SSL. Align DNS and system time. The link between the Frontend service and the API, and managed storage, use certificates issued by SmartAGT: no certificate is required for them.
The upstream firewall allows the workstation network to reach TCP 9000 on the VM. Data ports remain restricted to components. On the API server, TCP ports 8010 and 8011 stay available for the Internal listener and the local enrollment endpoint, both on loopback on a single VM. A reverse proxy can present the Frontend service on 443 in the project journey.
Ports and flows lists sources and destinations.
Project sizing
| Requirement | Resource to assess |
|---|---|
| Conversation load | Application CPU/RAM, provider concurrency, PostgreSQL connections and pools for concurrent conversations and long calls |
| Document ingestion | Volume throughput and capacity, vector indexes, RAG workers and embedding API for the corpus and concurrent ingestion |
| Local processing | Runtime CPU and memory, model cache and dependencies for OCR, local models or large pages |
| Generated files | Image and object size, retention and storage access |
| High availability | Role placement, shared data, PostgreSQL and storage availability |
| AI Gateway | Additional processes and pools, external client throughput |
Each ingestion worker can use several processes to read a native PDF. Under
smartagt.ingestion, an absent or zero ingestor_pool_size uses half the visible
logical processors, including affinity restrictions, with a minimum of 1 and a
maximum of 4. The 4 vCPU profile therefore uses 2 processes per worker. A positive
integer fixes this number; a negative value prevents startup. When several workers
share a host, assign an explicit budget to each. Keep CPU and memory headroom for
conversations and local models; local embedding and OCR inspection can still
dominate ingestion time.
This budget does not guarantee stable search latency when local models share the
same host: validate sizing under load.
As an indication, reading a 47 MB native PDF containing text and images reached about 0.50 GB of resident memory on Windows for one worker and its four processes combined. This measurement excludes embedding and OCR services and is not an upper bound: measure your corpus before sizing several workers.
The catalog's local OCR pipelines run on CPU. Their weights, decoded pages and concurrent calls require additional memory headroom and disk cache. Access to the weights repository is used for explicit preparation; extraction then reads the local cache, which must be accessible to the service account. Tesseract and its language data are not required.
These requirements are not implicitly included in the PoC's 16 GB. Project architectures explains their effect on component placement.