Vector Search Configuration

If you have an atlas vector search index defined for your collection you should be able to see a configuration of that index when you navigate to the Search Indexes tab in the collection settings:

Please note

You can not create an index in Florentine.ai. Once you create an index via Atlas or Mongo Shell you should see it on the Search Indexes tab as well.

Select Embeddings Model

When Florentine.ai creates a vector search it automatically triggers a call to the selected embedding model with the search query. This embedding model must be the same model the embedding field in the document was created with.

Based on the number of dimensions of the embedding field Florentine.ai preselects the most likely embedding model. However you should double check if its correct and you might need to change it:

Currently we support the most common embedding models from OpenAI and VoyageAI. Please contact us at [email protected] if you have embeddings created with another model, we'll be happy to integrate it.

Set Score Threshold

Every vector search in MongoDB returns a score that shows how closely a document matches the meaning of your query. This score ranges from 0 to 1, where 1 means a perfect match. In many cases, it's helpful to set a minimum score (score threshold) to filter out results with lower scores. This helps avoid showing results that are less relevant or confusing, and also reduces the number of documents returned.

Florentine.ai allows you to select one of three options for a score threshold, None, Dynamic and Static:

Setting the Score Threshold to None:

Returns all found documents that somehow semantically match, even when the similarity might be quite vague.

Setting the Score Threshold to Dynamic:

This is the default setting. Based on the question the AI selects a threshold on its own from in between the defined range (default range is the full scale between 0 and 1). You can define the range with the slider next to the select buttons:

I.e. if you set the range from 0.2 to 0.7 the AI will set the score threshold for a question between these two values. For very precise factual lookups it would set a higher threshold and for very broad or explanatory questions it would set a lower threshold.

Setting the Score Threshold to Static:

Sets the threshold to a fixed static value: