sweepx.agent¶
Offline-LLM natural-language control for the sweep stack — say
what you want in plain language and a local LLM turns it into a validated
sweep run. No cloud, no API keys.
"here is
vp_init.npyandobs.segy, run an FWI starting at 10 Hz" → a local LLM turns that into a validatedsweeptask and runs it.
Part of sweepx (pip install sweepx)
→ import sweepx.agent.
How it works¶
user (natural language + files)
│
▼
┌──────────────────────────┐ tool_call
│ Agent loop (agent.py) │ ───────────────► local LLM (OpenAI-compatible: vLLM / Ollama)
│ │ ◄─────────────── tool result (observation)
└───────────┬──────────────┘
│ dispatches to one of ~30 registered tools
▼
tools/ ── inspect_file · list_equations · make_synthetic_model · get_benchmark_model
· run_forward_sweep · build_fwi_spec · run_task · plot_* · ...
│
├─ discovery / modelling ──► sweep (core solver)
└─ build + execute + viz ──► sweepx.tasks.TaskRunner (production runner)
Tools import the geophysics stack lazily — a tool whose dependencies fail to
import returns a clear {"error": "… not importable"} instead of crashing, so the
agent always starts.
What the tools cover¶
pip install sweepx brings the solver and the task runner along, so the tools work
on a plain install; GIFs need pip install "sweepx[animate]" and SEG-Y files
pip install "sweepx[segy]" (the tools say so when the package is missing):
| tools | what they do |
|---|---|
inspect_file, check_parameters, make_synthetic_model, describe_task_schema |
discovery and inspection |
plot_*, compare_shot_gathers, list_equations |
plots, equation listing |
run_forward_sweep |
forward modelling (acoustic + elastic) |
list_benchmark_models / get_benchmark_model |
Marmousi / Overthrust |
build_*_spec, run_task, run_fwi, run_multiscale_fwi, … |
FWI / LSRTM through sweepx.tasks, the production runner |
-
Install, point it at a local LLM, run your first chat.
-
Tools and LLM backends (Ollama / vLLM / any OpenAI-compatible).
-
Prompts to start from.
-
The tool functions and the
BaseLLMbackend interface.