Interpreting Scores
Like all information retrieval systems, Vectara scores documents on how relevant they are to the query. It's important to understand the scoring system and how it changes based on the controls and query parameters you have provided.
Out of the box scores in Vectara:
- Can be either positive or negative
- Are larger/more positive as relevance is increased
- Are between -1 and 1 when not reranked
- Can be any real number -- positive or negative -- when reranked. However, scores when reranked are typically between about -10 and 10
See the sections below on "standard" and "reranked" results for details on how they differ and how to use them best.
Vectara provides an important control that can affect scores: custom dimensions, which allow you to boost or bury results based on metadata.
Comparison with keyword systems
If you've used a keyword-based system (BM25, TF/IDF, etc) and are used to the scoring mechanics, it's worth discussing the differences so you can understand what to expect with Vectara.
Limiting number of results
Unlike keyword systems which only match documents that exactly match the term(s) that have been searched, Vectara attempts to produce scores for the majority of documents in a corpus -- even those that have low relevance to the given query. For some use cases, it's desirable to have as many pages of results as possible, but for others, you may wish to apply logic to limit the number of results that are shown in the website or application.
In general, you can safely stop showing results with scores below 0.1 in all cases and below around 2 when using the reranker for most use cases. A more robust strategy for limiting results is to look for a sudden drop in scores. For example, if the scores are [0.7, 0.6, 0.55, 0.5, 0.2, 0.14], the large drop from 0.5 to 0.2 means results have gotten significantly worse, and can be used as a signal to stop returning results.
Comparing results across time and corpora
One of the advantages of Vectara's scoring system is that results are comparable across both different points in time and completely different corpora, even after you've indexed additional/different documents. This means you can split up your data across any corpora and not worry about how that splitting may affect scoring. This is because, unlike most keyword systems, Vectara does not rely on how relatively rare a term or set of terms is, but rather the absolute semantic significance of a document to a query.
Multilingual
Most keyword-derived search systems will instruct you to separate content for different languages into different corpora or indices because some terms are extremely common in one language but extremely rare in a different language. Splitting the indices helps ensure accurate term statistics. Vectara is different here, because it provides the absolute semantic significance scores. That is, Vectara isn't sensitive to situations like that where indexing a set of Spanish documents containing the word "hay" (a common word in Spanish) decreases the relevance of an English document in the same corpus that talks about the grain "hay" because Vectara takes the language and context into account by default.
Standard results response
By default, results from Vectara will be scored on a scale from -1 to 1, with 1 being a perfect match and -1 having absolutely nothing to do with the query. In practice, the vast majority will be scored between 0 and 1.
As with most search systems, there is no hard rule around when to cut off results as no longer very relevant. However, as a general rule, scores less than 0.1 tend to be of low quality and can typically be safely removed/ignored.
Reranked results response
Scores from Vectara after reranking are scored on a scale from -infinity to +infinity. Internally, the numbers returned from the reranker are derived from a logarithmic scoring system, which means that in practice, scores much higher than 10 or much lower than -10 should be rare.
As with standard results, there's no hard rule around when to cut off results, but scores above around 2.5 or so tend to be pretty good, though we advise users to test with their own data and some sample queries.