Skip to content

Hardware guide

Which server to buy, or rent, for Zettra (FSD §18.1, §21 Iteration 6). Zettra runs on one Linux server with Docker. How fast Ask answers depends almost entirely on the model server. Everything else is modest.

Pick a tier

Tier For CPU and memory GPU Disk Chat model Ask median
AI off Tickets, history and keyword search only 4 cores, 8 GB none 100 GB SSD none keyword results in under a second
Minimum A small team trying Ask 8+ cores, 32 GB none 250 GB SSD qwen3.5:4b 45–120 s (estimate)
Recommended Daily Ask for a team of up to 50 8+ cores, 32 GB 12–16 GB VRAM (e.g. RTX 4060 Ti 16 GB, RTX A4000) 250 GB SSD qwen3.5:9b 15 s or less (target)
Large team Heavy Ask use, many at once 16+ cores, 64 GB 24 GB+ VRAM (vLLM) 500 GB SSD Qwen3.5-27B or -35B-A3B 10 s or less (estimate)
  • Embeddings use bge-m3 (1.2 GB) on every AI tier.
  • The dev laptop (MacBook Air M4, 16 GB) runs qwen3.5:4b natively in about 18–21 s per answer. It is for development and demos, not for a team.
  • Bring your own key (BYOK) needs no GPU: the provider runs the model. Size the server like the AI-off tier.

Memory

PostgreSQL gets shared_buffers and effective_cache_size from .env:

Server memory PG_SHARED_BUFFERS PG_EFFECTIVE_CACHE_SIZE
8–16 GB 512MB 2GB
32 GB 4GB 16GB
64 GB 8GB 32GB

install.sh writes the 32 GB values on a host with 30 GB or more, and the small ones otherwise.

With local AI on a CPU-only server, the model shares this memory: qwen3.5:4b takes about 3.4 GB and bge-m3 about 0.7 GB.

Disk

At 100,000 tickets the database is about 8 GB with vectors (estimate, FSD §16): - half-precision vectors, about 2 GB; - the HNSW index, 2–3 GB; - text and other tables, about 3 GB.

Plan for: - attachments: whatever your teams upload (screenshots are typically 100–500 KB each); - backups: 14 daily dumps without vectors, plus a copy of every attachment; - models: 8 GB for the minimum tier, 12 GB for recommended.

Admin → System status warns when the attachments or backups disk reaches 80%.

Measured

The load test, deploy/loadtest/run.sh 100000, ran on the development container: 4 vCPU and 15 GB, with every service and k6 on one host and AI off. It uses 100,000 tickets (1,000,000 chunks), 50 users browsing and 5 asking, for 3 minutes:

Measure Target (FSD §18) Result
API p95 under 200 ms 142 ms
Ticket page p95 under 1 s 680 ms
Ask median, keyword path under 15 s 0.25 s
Failed requests under 1% 0%

Before Iteration 6, the API p95 was 313 ms and the page p95 1.03 s. Ask's keyword search ranked every chunk that held any word of the question, about 136,000 rows and 3 s of database time for a common word. It now ranks the chunks with all the words first, and at most 5,000 candidates per search.

The Ask figure leaves out the model: add the chat model's time from the tier table above, as measured by app eval. Run the test on your own server before a pilot. It seeds project LOAD in a scratch install and prints each measure against its target.

Network

  • Local or off AI: only ports 80 and 443 in. Nothing goes out after the install.
  • BYOK: outbound HTTPS to the provider only.
  • SSO (P1): outbound HTTPS to the identity provider.