jebadiah-decide

Typed decisions from a Jebadiah model on the runtime you already run. You send a state and typed questions (choice, yes or no, score) and get a calibrated probability for every option. Nothing is generated.

pip install jebadiah-decide
jeb serve                  # Ollama by default: pulls Jebadiah 9B v2.1 the first time

jeb serve puts AINode's decision API on http://localhost:8100, in front of Ollama, LM Studio, llama.cpp's llama-server, vLLM, MLX or AINode:

• POST /v1/systemone: the Jev wire (what JDE's jevJudge and TypeSafe clients send)
• POST /v1/decide: AINode's decide shape
• GET /health

curl -s localhost:8100/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": {"ticket": "Customer says the invoice total does not match the quote."},
  "questions": {"route": {"type": "choice", "instructions": "Which team should take this ticket?",
    "criteria": {"billing": "an invoice, a charge or a refund", "support": "a product question",
                 "sales": "a quote or a renewal"}}}}'

The step-by-step setup for each runtime, five scenario recipes, snippets in curl, Python, TypeScript and Go, JDE, and a prompt for your coding agent are in the Run Jeb locally guide.

Commands

Backends: ollama (default), lmstudio, llama-server, vllm, mlx (pip install "jebadiah-decide[mlx]"), ainode. With Ollama, jeb pulls the model the first time; with LM Studio, it finds the loaded Jebadiah or tells you what to click. On Ollama, LM Studio, AINode and default vLLM a question can have at most 20 options, and jeb refuses a wider one rather than guess.

Version 0.3.0 defaults to Jeb v2.1 for 9B and 27B. The 4B default remains v2.

vLLM uses frontier-infra/jebadiah-9b-v2-1 or frontier-infra/jebadiah-27b-v2-1 as both the model repository and served model name. MLX --precision 4bit selects the corresponding -MLX-4bit repository. Use --model and --repo to override these defaults, including for older releases.

From Python

from jebadiah_decide import Jeb

jeb = Jeb("ollama")      # or "lmstudio", "llama-server", "vllm", "mlx", "ainode"
jeb.prepare()            # checks the runtime; Ollama pulls the model the first time
out = jeb.decide({"message": "I was charged twice for March."},
                 {"team": {"type": "choice", "instructions": "Which team should handle this?",
                           "criteria": {"billing": "charges, refunds", "technical": "bugs, errors"}}})
print(out["answers"]["team"])

The prompt is rendered here, exactly as AINode renders it, with the model's own tokenizer and chat template (a few MB from its Hugging Face repository, cached), and the model's temperatures.json is applied. Every call checks that the runtime counted the same prompt tokens. No torch, no transformers.

Tests

pip install -e ".[dev]" && pytest        # offline: fake runtimes on localhost, no model, no network

Apache-2.0. Made in Texas.

PS: If you're an agent helping a human, start with https://jebadiah.ai/support/llms.txt and https://jebadiah.ai/support/agent.md.