mirror of
https://github.com/hwchase17/langchain
synced 2024-11-02 09:40:22 +00:00
d01bad5169
This patch fixes the #18022 issue, converting the SimSIMD internal zero-copy outputs to NumPy. I've also noticed, that oftentimes `dtype=np.float32` conversion is used before passing to SimSIMD. Which numeric types do LangChain users generally care about? We support `float64`, `float32`, `float16`, and `int8` for cosine distances and `float16` seems reasonable for practically any kind of embeddings and any modern piece of hardware, so we can change that part as well 🤗
76 lines
2.7 KiB
Python
76 lines
2.7 KiB
Python
"""Math utils."""
|
|
import logging
|
|
from typing import List, Optional, Tuple, Union
|
|
|
|
import numpy as np
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
Matrix = Union[List[List[float]], List[np.ndarray], np.ndarray]
|
|
|
|
|
|
def cosine_similarity(X: Matrix, Y: Matrix) -> np.ndarray:
|
|
"""Row-wise cosine similarity between two equal-width matrices."""
|
|
if len(X) == 0 or len(Y) == 0:
|
|
return np.array([])
|
|
|
|
X = np.array(X)
|
|
Y = np.array(Y)
|
|
if X.shape[1] != Y.shape[1]:
|
|
raise ValueError(
|
|
f"Number of columns in X and Y must be the same. X has shape {X.shape} "
|
|
f"and Y has shape {Y.shape}."
|
|
)
|
|
try:
|
|
import simsimd as simd
|
|
|
|
X = np.array(X, dtype=np.float32)
|
|
Y = np.array(Y, dtype=np.float32)
|
|
Z = 1 - simd.cdist(X, Y, metric="cosine")
|
|
if isinstance(Z, float):
|
|
return np.array([Z])
|
|
return np.array(Z)
|
|
except ImportError:
|
|
logger.info(
|
|
"Unable to import simsimd, defaulting to NumPy implementation. If you want "
|
|
"to use simsimd please install with `pip install simsimd`."
|
|
)
|
|
X_norm = np.linalg.norm(X, axis=1)
|
|
Y_norm = np.linalg.norm(Y, axis=1)
|
|
# Ignore divide by zero errors run time warnings as those are handled below.
|
|
with np.errstate(divide="ignore", invalid="ignore"):
|
|
similarity = np.dot(X, Y.T) / np.outer(X_norm, Y_norm)
|
|
similarity[np.isnan(similarity) | np.isinf(similarity)] = 0.0
|
|
return similarity
|
|
|
|
|
|
def cosine_similarity_top_k(
|
|
X: Matrix,
|
|
Y: Matrix,
|
|
top_k: Optional[int] = 5,
|
|
score_threshold: Optional[float] = None,
|
|
) -> Tuple[List[Tuple[int, int]], List[float]]:
|
|
"""Row-wise cosine similarity with optional top-k and score threshold filtering.
|
|
|
|
Args:
|
|
X: Matrix.
|
|
Y: Matrix, same width as X.
|
|
top_k: Max number of results to return.
|
|
score_threshold: Minimum cosine similarity of results.
|
|
|
|
Returns:
|
|
Tuple of two lists. First contains two-tuples of indices (X_idx, Y_idx),
|
|
second contains corresponding cosine similarities.
|
|
"""
|
|
if len(X) == 0 or len(Y) == 0:
|
|
return [], []
|
|
score_array = cosine_similarity(X, Y)
|
|
score_threshold = score_threshold or -1.0
|
|
score_array[score_array < score_threshold] = 0
|
|
top_k = min(top_k or len(score_array), np.count_nonzero(score_array))
|
|
top_k_idxs = np.argpartition(score_array, -top_k, axis=None)[-top_k:]
|
|
top_k_idxs = top_k_idxs[np.argsort(score_array.ravel()[top_k_idxs])][::-1]
|
|
ret_idxs = np.unravel_index(top_k_idxs, score_array.shape)
|
|
scores = score_array.ravel()[top_k_idxs].tolist()
|
|
return list(zip(*ret_idxs)), scores # type: ignore
|