Retrieval

Problem

Ranking a big volume of contents will have poor latency on the service.

Goal

Before ranking contents, have a fast model that can find relavant contents from a large content corpus.

How It's Done

High Level

1. Model vector representations of queries and contents in the same d-dimensional space.

2. For a given query vector use Approximate Nearest Neighbor (ANN) to retrieve relevant contents.

Offline Metric

Key goal of the retrieval model isn't to rank, but to find all the relavent contents corresponding to a given query. Therefore, we want to focus on recall rather than precision. Also, we want the retrieval stage to be diverse so coverage is also important.

For a given test set with $N$ queries we look find the top $k$ contents

• recall@k: $\frac{1}{N} \sum \frac{\text{True Positives}}{\text{True Positives + False Negatives}}$
• coverage@k: $\frac{\text{total number of distinct contents retrieved}}{\text{total number of contents in corpus}}$

Model

Two Tower Embedding Model

Learn two independent neural networks that encode query and content onto the same d-dimensional space.

Given an encoded query vector $q$ and content vector $c$, the model should learn such that the similarity between $q$ and $c$, $s(q, c)$ are high.