For now, when I check Similarity and Traffic, I'm having trouble understanding how they have been implemented in the schema and, therefore, how to evaluate them. Do you have a solution, @rngadam?

Similarity: Represents the primary signal in our scoring mechanism. It denotes the measurement of how closely a document aligns with the user's query, reflecting the relevance of the document to the search criteria. It is not a static precomputed score but a dynamic metric that is computed on the fly from the search results. This ensures that the relevance is tied to the specific query, going beyond simple keyword matching. Documents with higher similarity scores are considered more relevant. By prioritizing similarity as the first signal in our scoring process, we aim to deliver search results that are more accurate.

Traffic: The frequency with which users consult a document influences its score. Popular or frequently accessed documents are given higher scores with the help of web traffic logs, indicating their relevance and importance to users. Warning: The home page is rated really high since it's where every user land at first.

Similarity is dynamic, it cannot by definition be precomputed. we the 1 - cosine distance between chunk vector and query vector for this:

Traffic comes from the webserver logs. I'm imported and computed a score per page in the past already:

https://github.com/ai-cfia/ailab-db/blob/main/sql/compute-traffic-score.sql

After testing the score seems to be accurate.