FAISS

This backend uses Facebook's FAISS library for approximate nearest neighbour search. It is the recommended default for most datasets and ships with three presets.

---

When to use

• Use as the default choice for most low-to-medium dimensional datasets
• It performs best when datasets are large and an accuracy loss is worth the reduced computation
• It performs worst on very high-dimensional data

---

How it works

At fit time, FAISS builds an index over the validation features. The index type determines the tradeoff between speed and recall. flat performs exact search via FAISS's optimised L2 scan. ivf (Inverted File Index) partitions the dataset into cells using k-means, then at query time searches only a subset of cells determined by n_probes, where more probes means higher recall but slower queries. n_cells is set automatically to sqrt(n_samples) if not specified, capped at 4096.

---

Presets

Preset	index_type	n_probes	Recall
balanced	ivf	50	~98%
fast	ivf	30	~95%
turbo	flat	—	100%

---

Parameters

Parameter	Type	Default	Description
k	int	10	Number of neighbours
index_type	str	"flat"	Index type: "flat", "ivf", or "hnsw"
n_cells	int	sqrt(n_samples)	Number of IVF cells. Only used when index_type="ivf"
n_probes	int	50	Number of cells searched per query. Only used when index_type="ivf"
hnsw_M	int	32	HNSW graph connections per node. Only used when index_type="hnsw"
hnsw_efConstruction	int	400	HNSW build-time search width. Only used when index_type="hnsw"
hnsw_efSearch	int	200	HNSW query-time search width. Only used when index_type="hnsw"

---

Dependencies

faiss-cpu>=1.7

---

Example

from deskit.des.knorau import KNORAU

# Using a preset
router = KNORAU(task="classification", metric="accuracy", mode="max", k=20, preset="balanced")
router.fit(X_val, y_val, val_preds)
weights = router.predict(x)

# Custom IVF configuration
router = KNORAU(task="classification", metric="accuracy", mode="max", k=20,
                preset="custom", finder="faiss", index_type="ivf", n_probes=80)
router.fit(X_val, y_val, val_preds)
weights = router.predict(x)

---

Notes

FAISS IVF requires a minimum of 40 samples per cell to train. If the validation set is too small, the library will automatically reduce n_cells and emit a warning.

FAISS Flat may have floating-point precision issues on data with 2 or fewer feature; a warning will be emitted in that case.