HNSW Parameter Tuning

Core Concept: The Trade-off Triangle

Before tuning these parameters, it's crucial to understand that they collectively affect a core trade-off triangle:

• Recall: The accuracy of search results. Higher recall means less likely to miss true nearest neighbors.
• Query Speed/Latency: The time required to execute a search. Faster speed means lower latency.
• Resource Consumption: Primarily refers to the memory size occupied by the index and the time required to build the index.

Typically, you cannot simultaneously achieve the highest recall, fastest query speed, and lowest resource consumption. Tuning these parameters is about finding the balance point among the three that best suits your application scenario.

---

Parameter Details

1. m (Max Connections)

• What it is: This is the maximum number of connections (out-degree) that each node (vector) in the graph is allowed to have when building the graph. Think of it as a social network where each person can have at most this many "friends".
• What it does:
  • Determines graph "density": Higher m values create denser graphs with more connection options between nodes.
  • Affects index quality and memory: A denser graph typically has higher quality, providing better search paths and thus improving recall potential. However, more connections mean storing more edge information, leading to larger index files and higher memory usage.
• Performance impact:
  • Increasing m → Improves recall upper limit, but significantly increases memory usage and slightly increases index build time.
  • Decreasing m → Reduces memory usage, but may sacrifice recall, especially on complex or high-dimensional datasets.

2. ef_construction (EF for Construction)

• What it is: This parameter is used during index construction. When a new node is to be inserted into the graph, ef_construction defines the size of a dynamic list (candidate pool), from which the algorithm selects the best m neighbors to connect to the new node.
• What it does:
  • Determines index construction quality: Higher ef_construction means the algorithm "works harder" when finding neighbors for new nodes, searching more widely and finding higher quality connections. A high-quality graph is the foundation of high recall.
• Performance impact:
  • Increasing ef_construction → Higher index quality, helping improve final query recall. But significantly increases index construction time.
  • Decreasing ef_construction → Faster index construction, but may sacrifice index quality, requiring higher ef_search during subsequent queries to achieve the same recall.

3. ef_search (EF for Search)

• What it is: This parameter is used during search execution. Similar to ef_construction, it defines the size of a dynamic list (candidate pool) used during search. The algorithm starts from entry points, continuously explores the graph, and places "promising" neighbors into this candidate pool until the search ends.
• What it does:
  • Directly controls search depth and breadth: This is the "knob" that directly affects recall and speed during queries.
• Performance impact:
  • Increasing ef_search → Broader search scope, visiting more nodes, significantly improving recall. However, computation also increases, leading to increased query latency (slower speed).
  • Decreasing ef_search → Smaller search scope, faster query speed, but may miss true nearest neighbors, leading to decreased recall.

Important note: ef_search must be greater than or equal to k in the top-k results you want to return.

---

When to Modify?

Here are some typical application scenarios and corresponding tuning recommendations:

Summary and Tuning Recommendations

1. Start with default values: Don't randomly modify at the beginning. Use default values for baseline testing to understand your system's baseline performance.
2. Determine core business metrics: What can your application least tolerate? High latency, inaccurate results, or memory overflow? This determines your tuning direction.
3. Step-by-step adjustment and testing:
   • Step 1: Determine m. m primarily affects memory and cannot be changed once the index is built. Based on your memory budget and recall expectations, choose an appropriate m (typically 16-48 is a good range).
   • Step 2: Determine ef_construction. This value affects index building time. If you're not sensitive to index building time (e.g., one-time construction, infrequent updates), you can set it higher (e.g., 400+) for better index quality.
   • Step 3: Dynamically adjust ef_search. This is the most flexible parameter and can be specified for each query. Through continuous testing, find an optimal ef_search value that meets recall and latency requirements on your dataset.