TensorFlow.js · in your browser

Danamam

Teach a lightweight buddy from your own Q&A pairs: questions become sentence embeddings, then a small classifier picks the closest answer after quick on-device training.

Universal Sentence EncoderPrivate — no server

Model failed to load

Reload the page or check your connection.

How it works

A short walkthrough of the flow and the idea behind it — kept simple.

  1. 1

    Load the sentence encoder

    Universal Sentence Encoder runs in the browser and turns each question into a 512-dimensional vector that captures meaning.

  2. 2

    Collect question / answer pairs

    Each pair teaches the buddy: when a question is similar to this, return this answer. More examples improve separation between answers.

  3. 3

    Train a classifier head

    A small network (Dense + Dropout + Softmax) learns to map the question vector to one of the unique answers in your data — it selects an answer, it does not freely generate text.

  4. 4

    Chat and export

    After training you can ask questions and get the closest answer. Export saves weights, architecture, and the answer list to a JSON file.

More detail

Learning needs more than one class or enough examples to tie vectors to answers. With only one pair there is not enough structure to generalize.
If you reuse the same answer for several different questions, it counts as one class. The model output size equals the number of distinct answer strings you used.
The encoder is multilingual; quality depends on how you phrase training data. Use phrasings close to what end users will type.
No — load, train, and inference all run in the browser. Export only happens when you click download.

Train your buddy

Add a question and the answer you want tied to it. Repeat with different phrasings for the same intent to improve generalization.

Pair progress0 pair(s) · minimum 2
0%

Add at least two pairs to enable Start training.

No pairs yet

Enter a question and answer, then click Add pair.