Test-ticketing / app /embedding.py
jonathanjordan21's picture
Create app/embedding.py
8b12d82 verified
Raw History Blame Contribute Delete
2.48 kB
import os
import joblib
import numpy as np
from app.models import models
PCA_PATH = os.getenv(
"PCA_PATH",
"artifacts/embedding_pca.joblib"
)
def generate_embeddings(request):
model = models.embedding()
vectors = model.encode(
request.input,
convert_to_numpy=True,
normalize_embeddings=False,
show_progress_bar=False
).astype(np.float32)
# PHASE 1: Embedding generation
original_dimensions = vectors.shape[1]
# PHASE 2: Dimensionality reduction
if request.dimensions:
if request.dimensions > original_dimensions:
raise ValueError("Requested dimensions exceed model output")
if request.dimensions < original_dimensions:
if not os.path.exists(PCA_PATH):
raise ValueError(
"PCA artifact not configured. "
"Train and deploy PCA before reducing dimensions."
)
pca = joblib.load(PCA_PATH)
if pca.n_features_in_ != original_dimensions:
raise ValueError("PCA input dimensions mismatch")
if pca.n_components_ != request.dimensions:
raise ValueError("PCA output dimensions mismatch")
vectors = pca.transform(vectors)
# Optional normalization after projection
if request.normalize:
norms = np.linalg.norm(
vectors,
axis=1,
keepdims=True
)
vectors = vectors / np.maximum(norms, 1e-12)
# PHASE 3: Precision quantization
scales = None
if request.precision == "float32":
output = vectors.astype(np.float32).tolist()
elif request.precision == "float16":
output = vectors.astype(np.float16).tolist()
elif request.precision == "int8":
# Symmetric per-vector quantization
max_values = np.max(
np.abs(vectors),
axis=1,
keepdims=True
)
scale_array = np.maximum(
max_values / 127.0,
1e-12
)
quantized = np.clip(
np.round(vectors / scale_array),
-127,
127
).astype(np.int8)
output = quantized.tolist()
scales = scale_array.flatten().tolist()
return {
"dimensions": vectors.shape[1],
"precision": request.precision,
"embeddings": output,
"scales": scales,
"original_dimensions": original_dimensions
}