Skip to content
localcode

Chat and Decisions

OpenJev supports two ways to answer. Chat generates text in the usual conversation. Decisions returns a choice, a yes/no probability, or an ordered score about content you supply.

Choose a mode

Section titled “Choose a mode”

Load OpenJev from /models, then type /mode and choose Decisions. You can also use /decide or start with:

Terminal window
localcode --mode decisions

In the decision screen, enter a question and optional content. Choose an answer type:

Type Input Result
Yes / no A question such as “Was the customer charged twice?” Probabilities for yes and no
Choice Between 2 and 52 different labels, one per line A selected label and probabilities for every label
Score Between 2 and 10 levels, lowest to highest Probabilities for each level and a weighted score from zero to the last level’s index

Optionally enter a PNG, JPEG or WebP image path, up to 5 MB. Image decisions require OpenJev’s vision projector. Select Run decision to evaluate; Back to chat returns to the conversation you left. Chat is the default on each launch.

Decisions call /v1/systemone directly through the authenticated local supervisor. They do not send a chat prompt, execute coding tools, or take actions. The form and result stay in the current interface’s memory; switching back preserves your chat session. A model switch during evaluation invalidates its result.

The windows share the loaded model. Decisions currently require OpenJev; other models produce an explicit error. There is no automatic fallback to a text answer.

Use the CLI

Section titled “Use the CLI”

Keep LocalCode open with OpenJev loaded, then run:

Terminal window
localcode decide --question "Which team should handle this?" \
  --type choice --option billing --option delivery \
  --state "I was charged twice for the same order."

The command prints JSON with the model, answers and usage. --type yes-no is the default. Repeat --option for each choice or score level. Use --state-file PATH for UTF-8 content, or --image PATH for an image. Run localcode decide --help for the full options.

For multiple questions in one request, use the Local API.

Interpret the result

Section titled “Interpret the result”

Probabilities describe the model’s prediction. Check calibration and accuracy on your own examples before using them to route work; a high probability does not authorize an action. Score levels are ordinal, so a weighted score depends on the levels you supplied.

OpenJev’s weights use CC BY-NC 4.0. The catalog identifies their non-commercial license.