Repository navigation
Description
Have I written custom code (as opposed to using a stock example script provided in MediaPipe)
Yes
OS Platform and Distribution
macOS 27.0.1
Mobile device if the issue happens on mobile device
N/A
Browser and version if the issue happens on browser
Chromium 153.0.8010.12
Programming Language and version
JavaScript
MediaPipe version
@mediapipe/tasks-decision 1.2.2 (same in 1.2.0)
Bazel version
N/A
Solution
Decision Maker
Android Studio, NDK, SDK versions (if issue is related to building in Android environment)
N/A
Xcode & Tulsi version (if issue is related to building for iOS)
N/A
Describe the actual behavior
normalizePrior is declared and documented on BooleanQuestion and ChoiceQuestion, but evaluateBoolean() and evaluateChoice() never send it to the Wasm module. Setting it has no effect, and there's no error or warning.
The same option works through evaluate():
| Call | normalizePrior off |
normalizePrior: true |
|---|---|---|
evaluateBoolean(), probability of true |
0.4926 | 0.4926 |
evaluate(), boolean question, probability of true |
0.4926 | 0.1933 |
evaluateChoice(), pick |
exchange 0.5047 |
exchange 0.5047 |
evaluate(), choice question, pick |
exchange 0.5047 |
refund 0.5901 |
I used embeddinggemma-2-text-270m.litertlm with maxNumTokens: 256. With 1.2.0, I also ran 600 tasks through these two methods with the option on and off, and all 600 results were identical.
The cause seems to be in decision_maker.ts: neither method reads question.normalizePrior, and the Wasm signatures for evaluateBoolean and evaluateChoice have no parameter for it.
Describe the expected behaviour
normalizePrior: true should do the same in evaluateBoolean() and evaluateChoice() as it does in evaluate().
If these two methods aren't meant to support it, the option should be removed from BooleanQuestion and ChoiceQuestion, so that TypeScript flags it and people know to use evaluate().
Standalone code/steps you may have used to try to get what you need
import { DecisionMaker, FilesetResolver } from '@mediapipe/tasks-decision';
const wasm = await FilesetResolver.forDecisionTasks(
'https://cdn.jsdelivr.net/npm/@mediapipe/[email protected]/wasm',
);
const maker = await DecisionMaker.createFromOptions( wasm, {
baseOptions: {
modelAssetPath: 'https://huggingface.co/litert-community/embeddinggemma-2-text-270m-litert-lm'
+ '/resolve/main/embeddinggemma-2-text-270m.litertlm',
},
maxNumTokens: 256,
} );
const text = 'You charged my card twice!';
const condition = 'This customer is upset.';
const off = await maker.evaluateBoolean( text, { condition } );
const on = await maker.evaluateBoolean( text, { condition, normalizePrior: true } );
console.log( off.probabilityTrue, on.probabilityTrue ); // 0.4926 0.4926
const { upset } = await maker.evaluate( text, {
questions: [ { id: 'upset', type: 'boolean', prompt: condition, normalizePrior: true } ],
} );
console.log( upset.probability ); // 0.1933Other info / Complete Logs
Workaround: call evaluate() with normalizePrior: true on the question.
A comment on #6384 mentions this in passing, but I didn't find an issue for it.
From reading decision_maker.ts, BooleanQuestion.options and ChoiceQuestion.scoringMode aren't read either. I can't confirm those two from behavior, because they made no difference on this model, and options made none through evaluate() either.
Activity
akashvverma1995 commentedon Nov 10, 2026
Hi @agustinbarrientos,
Thanks for the detailed report. The reproduction steps and workaround are super helpful. I will reproduce it from my end. I'll look into it and get back to you!
akashvverma1995 commentedon Nov 12, 2026
Hi @agustinbarrientos
NormalizePrior forwarding for evaluateBoolean(), evaluateChoice(), and their batch variants has been added directly to the Wasm bindings and Web SDK in commit 0528108, and is now available in @mediapipe/[email protected]. Please upgrade to 1.2.3 and let us know if everything works as expected!