open-inference-spec

Prediction

A prediction is a column that captures the output of the model. Predictions are prefixed with the category :prediction: in the column name.

Single Value Predictions

predictions have a few reserved specifiers that can be used to capture specific reserved information about the prediction. The following is a list of specifiers that are supported by OpenInference:

Here is an example of a set of prediction columns:

</tbody> </table> ## Text Predictions In some cases (e.g. LLMs, NLP), the prediction is a sentence or paragraph. In this case, the prediction is prefixed with `:prediction:text:`. Here is an example of a text prediction:
:prediction.float.score:confidence :prediction.str.label:predicted_animal
0.8 cat
0.9 dog
0.2 cat
</tbody> </table> ## Embedding Predictions In some cases (e.g. LLMs, NLP), the prediction can be encoded as an embedding. The following is an example of a set of embedding predictions:
:prediction.text:response
Yes, the capital of Japan is Tokyo
The population of Ireland is less than England
the quick brown fox jumped over the lazy dog
</tbody> </table> Embedding predictions can also be associated with a URL to the raw data of the embedding. This is a string but is a URL to a file (e.g., an image or audio file). In this case, the prediction is prefixed with `:prediction.url:`. Here is an example of a set of embedding predictions with a URL:
:prediction.vector:response :prediction.text:response
[0.1, 0.2, 0.3] Yes, the capital of Japan is Tokyo
[0.1, 0.2, 0.3] The population of Ireland is less than England
[0.1, 0.2, 0.3] the quick brown fox jumped over the lazy dog
</tbody> </table>
:prediction.vector:response :prediction.text:response :prediction.url:response
[0.1, 0.2, 0.3] An image of tokyo https://example.com/image.jpg
[0.1, 0.2, 0.3] A map of Ireland https://example.com/image.jpg
[0.1, 0.2, 0.3] The Eiffel tower https://example.com/image.jpg