If vector search is missing many true neighbors, improving recall typically involves adjusting index parameters, refining the search process, or adding post-processing steps like re-ranking. Here are three effective techniques:
1. Adjust Index Parameters for Broader Exploration
Vector indexes like HNSW (Hierarchical Navigable Small World) or IVF (Inverted File) rely on parameters that balance speed and recall. For example, in HNSW, increasing efSearch (the number of candidate nodes explored during search) allows the algorithm to consider more potential neighbors, improving recall at the cost of slower queries. Similarly, in IVF, raising nprobe (the number of clusters searched) expands the scope of the search. If your index uses 100 clusters with nprobe=10, increasing it to nprobe=20 forces the system to scan twice as many clusters, making it more likely to find true matches. However, these adjustments increase latency, so testing incremental changes is critical to find a practical balance.
2. Re-Rank with Exact Search on Initial Results
Approximate nearest neighbor (ANN) indexes prioritize speed but sacrifice precision. A hybrid approach combines ANN with a secondary exact search on the initial results. For instance, first retrieve 200 candidates using an approximate index, then re-rank those 200 using brute-force exact search to pick the top 10. This leverages the speed of ANN for broad exploration while ensuring the final results are accurate. Libraries like FAISS support this via a IndexFlat layer on top of an IVF or HNSW index. While re-ranking adds computational overhead, it’s manageable if the initial candidate pool is kept small (e.g., 100–200 items). This method is particularly effective when the ANN index’s recall is inconsistent but fast enough to generate a reasonable candidate subset.
3. Preprocess Data and Validate Embeddings Poor recall can stem from suboptimal embeddings or misconfigured distance metrics. For example, if vectors aren’t normalized when using cosine similarity, similarity calculations will be inaccurate. Ensure embeddings are normalized (scaled to unit length) and that the distance metric aligns with the data’s characteristics—Euclidean distance for L2-normalized vectors, cosine for angular similarity. Dimensionality reduction (e.g., PCA) can also help by removing noise from high-dimensional embeddings. Additionally, test your embeddings with a brute-force exact search to establish a recall baseline. If brute-force performs well but ANN does not, the issue lies in the index setup rather than the data. If both underperform, revisit the embedding model or training data.