Datathon Project: config builder

[forsyth2]

This PR should NOT be merged. It is to show a prototype developed for the January 2026 LLNL Datathon. See here for an earlier (unrelated) zppy "hackathon" project.

Goal: a user should be able to give a simulation data directory path and get back a zppy cfg that if not perfect, at least gives a really good starting point.

Motivating need: the zppy cfg is quite complex, with a very large number of parameters. It would therefore be useful to create custom starter cfg files for users to build on.

Architecture: This implementation consists of 3 layers, building on top of each other.

1. A script to extract available data from a simulation output directory. That's simulation_output_reviewer.py.
2. A script to generate a starting point cfg file based on the data found. That's zppy_config_generator.py.
3. An agentic AI question-answering system that would allow users to customize that cfg (the "answer" would be a better cfg than the starting point). Examples: a user could request "diagnostics relating to this particular physical phenomena" or "diagnostics on a hindcast during these particular years"

Layer 3 is more exploratory and the core of the Datathon challenege. Layers 1-2 will likely be cleaned up and merged into zppy as a distinct PR at a later date.

I've gone with this approach because a lot of the zppy cfg can in fact be constructed in a deterministic manner, lending itself more to a script than an AI agent. However, the more natural-language oriented layer 3 is a better fit for an AI agent, and the agent would also have a more-informed starting point rather than needing to "learn" all the rules of our data structuring.

[forsyth2]

Current status (end of datathon day 1): Layer 1 is complete. Layer 2 is in-progress, nearing completion.

Layer 2 action items:

• Some subtasks are missing even though I know the data is available. It seems likely to be from a bug in the Dependency logic.
• Find a way to infer the various data paths
• Find a better way to determine/suggest year increments.

Layer 3 action items:

• Add ollama to environment and begin experimenting with it to construct the question-answering system.

[forsyth2]

Layer 2 action items above implemented by second commit

[forsyth2]

The third commit is an initial implementation of an agentic AI workflow. run_agent.py was largely generated with Claude, with fixes added by me. I will now proceed with actually installing ollama and testing it out.

Notes from Claude on Ollama:

Usage Limits - Completely Unlimited! 🎉
Ollama is 100% free and open-source with NO limits:

✅ No token limits - Process as much as you want
✅ No API costs - Runs locally on your machine
✅ No internet required - After downloading models, works offline
✅ No rate limits - Run as many queries as you need
✅ Privacy - Your data never leaves your machine

Requirements:

Disk space: ~4.7GB for llama3.1:8b, ~40GB for llama3.1:70b
RAM: 8GB minimum for 8B model, 64GB for 70B model
GPU: Optional but highly recommended (10x+ faster with GPU)

Recommended for your use case:

Start with llama3.1:8b - fast, runs on most machines
If accuracy isn't perfect, try llama3.1:70b (needs more resources)
Or try llama3.2:3b if you have limited RAM

[forsyth2]

Following directions from @tomvothecoder in https://github.com/aims-group/llnl-datathon-2026/pull/1/files:

### 3. Option B: Local LLM via Ollama (Optional)

Ollama is used as a **local LLM service** for agentic workflows.

> **Important:** Ollama is **not managed by Conda**.
> It must be installed at the system level.
Install Ollama by following the official instructions:
[https://ollama.com/download](https://ollama.com/download)