(curious) How are the threshold values being determined?

[szelok]

https://github.com/aurelio-labs/semantic-router/blob/main/semantic_router/layer.py#L153C1-L158C40

      if isinstance(encoder, OpenAIEncoder):
            self.score_threshold = 0.82
        elif isinstance(encoder, CohereEncoder):
            self.score_threshold = 0.3
        else:
            self.score_threshold = 0.82

[jamescalam]

depends on the route layer, but the default RouteLayer is using semantic similarity followed by aggregation and filtering using the score_threshold values you showed above

[szelok]

@jamescalam - my question was how we should determine the threshold score. For example, you set the threshold to 0.82 for OpenAI but 0.3 for Cohere (a big difference). Any advice?

Also, it feels like it requires good 'utterance' engineering. For example I tried asking 'What do you think about Joe Biden' in your example (fro HuggingFaceEncoder). It didn't match any routes. Just my two cents.