Sanad setup

Sanad (سند) is Nassila's passage-grounding worker. It compares sentences in your draft to source excerpts using an OpenAI-compatible LLM endpoint you configure — local by default, cloud optional.

Model downloads

Sanad GGUF files are not bundled in the installer. Download from Hugging Face (bring your own file):

Model	Model ID	VRAM (approx.)	Hugging Face
Sanad 9B (sole tier)	`nassila-sanad-9b`	Q2_K ~3.6 GB → Q8_0 ~8.9 GB (Q6_K ~6.9 GB)	nassila-sanad-9b

Six GGUFs (Q2_K / Q3_K_M / Q4_K_M / Q5_K_M / Q6_K / Q8_0) share the same trained weights — pick the largest your memory allows.

In Nassila: Settings → Passage grounding — the Sanad bar shows a single 9B tier; the model field auto-fills nassila-sanad-9b when using a local Nassila GGUF.

Set up Sanad dialog with the Sanad 9B download link and privacy note — The first-use Set up Sanad dialog — grounding sends passage text only to the endpoint you configure.

Configure in Nassila

Open Passage grounding settings
Settings → Passage grounding (or the in-app Set up Sanad link — full detail lives here on the website).
Choose local or cloud
Under Provider, pick a local runner (LM Studio, Ollama, vLLM, Custom) for the Nassila Sanad GGUF, or Cloud API for an experimental third-party model.
Set base URL and model ID
Base URL must match your runner or API. Model ID must match what the server exposes (ollama list, LM Studio loaded model, or cloud model name).
Test connection
Use Test connection in settings before your first audit. Fix base URL, model id, or API key until it passes.
Enable in Manuscript
Toggle Sanad on the Sanad bar when running a manuscript audit.

Runner and cloud setup

Pick the tab that matches your provider. Local tabs use the Nassila-trained Sanad model; Cloud API uses whatever model your vendor exposes (quality not guaranteed).

Install LM Studio
Download from lmstudio.ai.
Load a Sanad GGUF
Download nassila-sanad-9b from Hugging Face (table above) and load the file in LM Studio.
Start Local Server
In LM Studio, start the Local Server (OpenAI-compatible API on port 1234 by default).
4
Match Nassila settings
Preset LM Studio. Base URL:
```
http://localhost:1234
```
Model field: nassila-sanad-9b or the exact id LM Studio shows for the loaded GGUF.

No template setup needed — Qwen3.5 thinking traces are stripped automatically in Nassila.

Privacy

Local runners — passage text goes to your machine only (localhost or your LAN server).
Cloud API — passage text goes to the vendor you configure.
Nassila never receives manuscript text for grounding; L1/L2 registry checks and OA fetch are separate actions you trigger.

Troubleshooting

Problem	What to try
Connection refused	Confirm runner server is running; check base URL port
Model not found	Match model id to LM Studio loaded model or `ollama list`
Slow or OOM	Use a smaller quant (Q2_K / Q4_K_M); close other GPU apps
Empty grounding	Enable Sanad toggle; confirm source excerpt is not `Unavailable`
Cloud test fails	Verify API key, `/v1` suffix on base URL, and model name with provider

More: Troubleshooting · Local models