The main type is DCEEConfig. Pass it to DCEEEngine(cfg) before build(). For every field, method signature, and copy-paste examples, see the API reference.
| Field | Role |
|---|---|
dim | Embedding dimension (must match your matrix columns). |
n_clusters | k-means clusters — routing granularity vs build cost. |
keyframe_every | Reset full vectors periodically inside a cluster to limit delta drift (vs compression). |
quantization | int8, float16, or float32 for stored deltas. |
n_probe / n_probe_max | How many clusters to score by keyframes; AMP can extend toward n_probe_maxwhen scores are ambiguous. |
adaptive_probe / adaptive_probe_margin | Enable AMP and set score-gap threshold for extra clusters. |
top_k_refine | How many coarse candidates get full-precision refinement before top-k. |
verbose | Logging and tqdm during build / tensor prep. |
tuned_for(n_vectors, dim)Static helper that returns a config with scale-aware heuristics (cluster count ~√N, probe budgets, refinement pool size). Override individual fields after calling it.
Higher n_probe / n_probe_max / top_k_refine usually improves recall at the cost of latency. Use tune_dcee.py in the repository for automated search over tradeoffs.