Modelfile

# Builds the `gemma-worker` Ollama model from a 4B Gemma base.
#
# gemma3:4b is the vision-capable 4B dense model, ~3.3GB at Q4_K_M -- comfortable
# on an 8GB card with room for the KV cache. To use a bigger model (needs the
# VRAM), change the FROM line and re-run:  ollama create gemma-worker -f Modelfile
#   - gemma3:12b   ~8GB at Q4  (tight on 8GB; good on 12GB+)
#   - or any newer Gemma tag your Ollama has  (ollama list / ollama pull)
FROM gemma3:4b

# 0.7 is a sane general default; the worker CLI overrides it per-call (0.0 for
# deterministic extraction). num_ctx 8192 fits the card; raise if you have VRAM.
PARAMETER temperature 0.7
PARAMETER num_ctx 8192

SYSTEM """
You are a helpful, concise local assistant. Be genuinely useful and direct.

- Answer plainly. No filler, no preamble, no restating the question.
- When a fact matters, say whether you actually know it or are inferring. Flag
  guesses as guesses; do not invent specifics you are unsure of.
- For extraction/classification tasks, output exactly what is asked and nothing
  else.
- You have no memory between calls and no access to any outside conversation, so
  work only from what is in front of you.
"""