Server presets

Shipped presets live in config/server_profiles.yml. User additions go to ~/.config/document-anonymizer/server.yml. Per-project overrides go to <output_dir>/.anon/server.yml. Resolution order: builtin ➜ user ➜ project (last wins).

Built-in catalog

Preset	Model	Quality	Ctx	Parallel	Disk	VRAM
default ★	Qwen 3.5 4B Claude-Opus distill Q4_K_M (CPU)	78	12 K	1	2.5 GB	~4.8 GB (RAM)
ministral-3-8b-reasoning-bf16	Ministral 3 8B Reasoning BF16	83	16 K	2	16.0 GB	~18.9 GB
rtila-qwen3.5-9b-q4	rtila Assistant Lite · 9B Q4_K_M	82	12 K	1	5.2 GB	~7.1 GB
jackrong-qwen3.5-4b-distill-q4 ★	Qwen 3.5 4B Claude-Opus distill Q4_K_M (GPU)	78	12 K	1	2.5 GB	~4.8 GB
qwen3.5-9b-bf16	Qwen 3.5 9B BF16	78	16 K	1	18.4 GB	~18.0 GB
ministral-3-8b-reasoning-q5	Ministral 3 8B Reasoning Q5_K_M	76	16 K	2	5.8 GB	~9.2 GB

Quality column = F1 × 100 from the 5-PDF pentest corpus. default and jackrong-qwen3.5-4b-distill-q4 share the same GGUF file; the only difference is n_gpu_layers (0 vs 99). Same model, two runtime configs.

How to choose

Hardware / goal	Pick
18+ GB VRAM, max quality	ministral-3-8b-reasoning-bf16
18+ GB VRAM, fewest false alarms	qwen3.5-9b-bf16
~7 GB VRAM, near-leader quality	rtila-qwen3.5-9b-q4
~6 GB VRAM, smallest "good" model ★ recommended	jackrong-qwen3.5-4b-distill-q4
~10 GB VRAM, reasoning quality	ministral-3-8b-reasoning-q5
No GPU	default (Jackrong 4B distill Q4 on CPU)

To pin a default for repeatable runs use ★ Set as default in the preset gallery, it persists to ~/.config/document-anonymizer/preferences.yml.

Server lifecycle (sidebar dot + auto-start)

The sidebar Server entry shows a small status dot:

• 🔴 red = offline
• 🟡 amber = starting (worker spawning, health endpoint warming up)
• 🟢 green = online

Click the dot to bring the active preset up. A pre-flight check runs first; on failure you're routed to the Server tab with a specific error dialog (missing binary, model not on disk, docker not in PATH, external server unreachable, etc.). When the server is already up or in flight the click just opens the Server tab.

The Server panel also exposes a persistent Auto-start server on launch toggle. Stored in ~/.config/document-anonymizer/app_settings.yml. Pre-flight applies here too (silent skip when the active preset can't start, no startup-time error dialog).

Customising

1. Open the Server view → Customize on any card.
2. Edit any field (cmd-line preview updates live).
3. Save → choose user or project scope.

Common tweaks:

• More parallel slots: parallel: 4 cuts wall time on long scans, at the cost of ctx_size / parallel per slot. Use the pre-flight token check (Pipeline → Run) to verify the slot still fits a ~7 500-token request.
• Bigger ctx: bump ctx_size to 32 K if you increase the detector's max_chunk_chars past 5 000 (rare).
• Different binary: point binary at your own llama.cpp build.

Quantisation policy

We ship both BF16 and quantised presets when the quant clears the quality bar. See BENCHMARKS.md for same-model BF16 vs. Q5_K_M / Q4_K_XL / Q8_K_XL deltas. KV cache stays at f16 across the board.