Monitoring Embedding Drift in Production Scikit-LLM Pipelines

In this article, you will learn what embedding drift is, why it matters for production large language models, and how to implement two practical techniques to detect it.

Topics we will cover include:

• The key approaches for detecting embedding drift in production machine learning systems, including model-based detection, centroid distance, and dimensionality reduction combined with statistical tests.
• How to implement a domain classifier and a centroid distance method using scikit-learn on simulated 384-dimensional embeddings.
• How to apply these same drift detection techniques to real text embeddings generated with a SentenceTransformer model via Scikit-LLM.

Introduction #

When a large language model (LLM) hits production, the story is far from over. User behavior inevitably evolves in the real world, and so does the data consumed by the model, typically encoded into numerical text representations called embeddings for its internal processing.

Therefore, it is crucial to track so-called embedding drifts to ascertain when a deployed model needs an update. However, traditional drift detection metrics designed for tabular data often fail when applied to high-dimensional embeddings.

This article starts by providing a brief outline of top techniques for detecting embedding drift, followed by an illustrative implementation of two of them, both simulation-based and in conjunction with the Scikit-LLM library for embedding generation.

Techniques for Effective Embedding Drift Detection #

Below we list three key approaches for accurately identifying embedding drift that have been remarkably put into practice in production LLMs:

• Model-based detection: This consists of training a domain-specific classifier, usually a binary classifier that has learned to distinguish between baseline data and new (drifted) production data. A model capable of easily telling them apart will be able to signal drifts when they occur.
• Centroid distance : Following classical anomaly detection algorithms, this strategy boils down to calculating the distance (often cosine for embedding data) between the center of mass of your baseline embedding vectors and that of new, incoming embedding vectors.
• Combining dimensionality reduction and statistical tests: This method entails compressing the embeddings to a lower dimension using UMAP or PCA, after which we apply standard drift tests such as Kolmogorov-Smirnov.

Interested in exploring further how they work? Let’s examine how to implement the core logic behind two of these techniques based on an open-source stack.

Illustrating Drift Detection on Simulated Embeddings #

Let’s build a mathematical foundation for two of the listed techniques using standard scikit-learn and simulated embeddings first. We generate an initial, random set of embeddings, after which we create another synthetic set — this time containing “production embeddings” that shift from the original embeddings’ mean to simulate the existence of data drift.

import numpy as np

n_samples = 500
n_features = 384

np.random.seed(42)
X_reference = np.random.normal(loc=0.0, scale=1.0, size=(n_samples, n_features))

X_production = np.random.normal(loc=0.3, scale=1.0, size=(n_samples, n_features))

1234567891011121314

import numpy as np # Simulating 384-dimensional embeddings (e.g. standard sentence-transformers output)n_samples = 500n_features = 384 # 1. Referencing Embeddings (Baseline / Training Data)# Imagine this is the data your LLM/Vector DB was originally populated withnp.random.seed(42)X_reference = np.random.normal(loc=0.0, scale=1.0, size=(n_samples, n_features)) # 2. Production Embeddings (New Data)# The original mean is shifted to loc=0.3 to simulate data drift (e.g. new topic emerging)X_production = np.random.normal(loc=0.3, scale=1.0, size=(n_samples, n_features))

Next, we train a domain classifier based on random forests to separate baseline data (labeled 0) from new, production data (labeled 1). If the accuracy metric — for instance, ROC-AUC — signals a high value, e.g. above 0.65, the classifier will trigger a drift alert.

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score

y_reference = np.zeros(n_samples)
y_production = np.ones(n_samples)

X_combined = np.vstack((X_reference, X_production))
y_combined = np.hstack((y_reference, y_production))

X_train, X_test, y_train, y_test = train_test_split(
    X_combined, y_combined, test_size=0.3, random_state=42
)

drift_classifier = RandomForestClassifier(n_estimators=50, max_depth=5, random_state=42)
drift_classifier.fit(X_train, y_train)

y_pred_proba = drift_classifier.predict_proba(X_test)[:, 1]
roc_auc = roc_auc_score(y_test, y_pred_proba)

print(f"Domain Classifier ROC-AUC Score: {roc_auc:.3f}")

if roc_auc > 0.65:
    print("ALERT: Significant embedding drift detected! Trigger retraining/review pipeline.")
else:
    print("System stable: Distributions are sufficiently similar.")

12345678910111213141516171819202122232425262728293031323334

from sklearn.ensemble import RandomForestClassifierfrom sklearn.model_selection import train_test_splitfrom sklearn.metrics import roc_auc_score # 1. Assigning labels: 0 for reference, baseline embeddings; 1 for production embeddingsy_reference = np.zeros(n_samples)y_production = np.ones(n_samples) # 2. Combining into a single datasetX_combined = np.vstack((X_reference, X_production))y_combined = np.hstack((y_reference, y_production)) # 3. Randomly splitting into train and test sets for the drift detectorX_train, X_test, y_train, y_test = train_test_split(    X_combined, y_combined, test_size=0.3, random_state=42) # 4. Training a lightweight Random Forest classifierdrift_classifier = RandomForestClassifier(n_estimators=50, max_depth=5, random_state=42)drift_classifier.fit(X_train, y_train) # 5. Evaluating the classifier using ROC-AUCy_pred_proba = drift_classifier.predict_proba(X_test)[:, 1]roc_auc = roc_auc_score(y_test, y_pred_proba) print(f"Domain Classifier ROC-AUC Score: {roc_auc:.3f}") # 6. Alerting Logic# If the metric score is around 0.5 it means the model can't tell the datasets apart (no drift detected).# Meanwhile, a score closer to 1.0 means they are easily distinguishable (high drift).if roc_auc > 0.65:    print("ALERT: Significant embedding drift detected! Trigger retraining/review pipeline.")else:    print("System stable: Distributions are sufficiently similar.")

Output:

Domain Classifier ROC-AUC Score: 0.970
ALERT: Significant embedding drift detected! Trigger retraining/review pipeline.

12

Domain Classifier ROC-AUC Score: 0.970ALERT: Significant embedding drift detected! Trigger retraining/review pipeline.

Alternatively, we can resort to the centroid calculation technique, also known as the “center of mass” method, measuring the distance between two centroids: one stemming from the baseline embeddings and one associated with the new, production embeddings. This method is computationally cheaper than the classifier method, but it incurs a loss of nuance (valuable information): after all, aggregating high-dimensional vectors into a single central point throws away complex distribution shapes, masking important patterns like multi-modal shifts or structural changes in the data.

from sklearn.metrics.pairwise import cosine_distances

centroid_ref = np.mean(X_reference, axis=0).reshape(1, -1)
centroid_prod = np.mean(X_production, axis=0).reshape(1, -1)

distance = cosine_distances(centroid_ref, centroid_prod)[0][0]

print(f"Centroid Cosine Distance: {distance:.4f}")

threshold = 0.05 
if distance > threshold:
    print("ALERT: Centroid distance exceeded threshold! System drifting.")
else:
    print("System stable: Centroids are aligned.")

1234567891011121314151617181920

from sklearn.metrics.pairwise import cosine_distances # 1. Calculating the centroid (mean vector) for both batches# axis=0 calculates the mean across all samples, resulting in a single 384-d vectorcentroid_ref = np.mean(X_reference, axis=0).reshape(1, -1)centroid_prod = np.mean(X_production, axis=0).reshape(1, -1) # 2. Calculating the distance (1 - Cosine Similarity) between the two centroids# A distance of 0 means identical direction; higher means they are drifting apartdistance = cosine_distances(centroid_ref, centroid_prod)[0][0] print(f"Centroid Cosine Distance: {distance:.4f}") # 3. Alerting Logic# Determining the exact threshold requires tuning in accordance with your specific model and baseline variancethreshold = 0.05 if distance > threshold:    print("ALERT: Centroid distance exceeded threshold! System drifting.")else:    print("System stable: Centroids are aligned.")

Output:

Centroid Cosine Distance: 0.9811
ALERT: Centroid distance exceeded threshold! System drifting.

12

Centroid Cosine Distance: 0.9811ALERT: Centroid distance exceeded threshold! System drifting.

No doubt the cosine distance value looks a bit exaggerated, due to a combination of the orthogonal nature of the distance metric used and the fact that the baseline data were generated randomly. A more realistic dataset would normally yield high distances in the presence of topic-driven data drifts, but not so extreme in the majority of cases. Let’s find out with a final example that uses Scikit-LLM to generate embeddings from real text.

Drift Detection on Generated Embeddings with Scikit-LLM #

The last code example uses Scikit-LLM as a wrapper for a Groq LLM specialized in embedding generation. It has been run on Google Colab, with an API key obtained from Groq (a free LLM repository) and stored in the “My Secrets” section of the left-hand side menu.

from sentence_transformers import SentenceTransformer
from google.colab import userdata
from skllm.config import SKLLMConfig

groq_api_key = userdata.get('GROQ_API_KEY')

SKLLMConfig.set_openai_key(groq_api_key)
SKLLMConfig.set_gpt_url("https://api.groq.com/openai/v1/")

vectorizer = SentenceTransformer('all-MiniLM-L6-v2')

texts_reference = [
    "How do I reset my password?",
    "Where is the billing menu?"
] * 100  # We multiply to simulate a larger dataset

texts_production = [
    "The new cryptocurrency system is failing",
    "How to mint an NFT on the platform?"
] * 100

X_reference = vectorizer.encode(texts_reference)
X_production = vectorizer.encode(texts_production)

y_reference = np.zeros(len(X_reference))
y_production = np.ones(len(X_production))

X_combined = np.vstack((X_reference, X_production))
y_combined = np.hstack((y_reference, y_production))

X_train, X_test, y_train, y_test = train_test_split(
    X_combined, y_combined, test_size=0.3, random_state=42
)
clf = RandomForestClassifier(n_estimators=50, max_depth=5).fit(X_train, y_train)

roc_auc = roc_auc_score(y_test, clf.predict_proba(X_test)[:, 1])

print(f"ROC-AUC Score: {roc_auc:.3f}")
if roc_auc > 0.65:
    print("DRIFT DETECTED! User queries have changed topic.")
else:
    print("System stable: Embeddings are consistent.")

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253

from sentence_transformers import SentenceTransformerfrom google.colab import userdatafrom skllm.config import SKLLMConfig # Securely extract the Groq API Key you may have previously stored in Colab secretsgroq_api_key = userdata.get('GROQ_API_KEY') # Redirecting scikit-LLM to Groq using API compatibility:SKLLMConfig.set_openai_key(groq_api_key)SKLLMConfig.set_gpt_url("https://api.groq.com/openai/v1/") # Since Groq does not have an embeddings API, we can use a free and very lightweight local modelvectorizer = SentenceTransformer('all-MiniLM-L6-v2') # Baseline raw texts and production texts, clearly with a drastic topic shifttexts_reference = [    "How do I reset my password?",    "Where is the billing menu?"] * 100  # We multiply to simulate a larger dataset texts_production = [    "The new cryptocurrency system is failing",    "How to mint an NFT on the platform?"] * 100 # Converting text to embeddingsX_reference = vectorizer.encode(texts_reference)X_production = vectorizer.encode(texts_production)  # Implementing Embedding Drift Detection Logic# Assign labels: 0 for reference, 1 for productiony_reference = np.zeros(len(X_reference))y_production = np.ones(len(X_production)) # Combining datasetsX_combined = np.vstack((X_reference, X_production))y_combined = np.hstack((y_reference, y_production)) # Training the domain classifierX_train, X_test, y_train, y_test = train_test_split(    X_combined, y_combined, test_size=0.3, random_state=42)clf = RandomForestClassifier(n_estimators=50, max_depth=5).fit(X_train, y_train) # Calculating drift using ROC-AUCroc_auc = roc_auc_score(y_test, clf.predict_proba(X_test)[:, 1]) print(f"ROC-AUC Score: {roc_auc:.3f}")if roc_auc > 0.65:    print("DRIFT DETECTED! User queries have changed topic.")else:    print("System stable: Embeddings are consistent.")

The process is similar to what we saw earlier. The main difference lies in the data used, which are now embeddings generated from real text examples. Due to the intentionally drastic topic difference between the two datasets, the classifier can perfectly distinguish between baseline and production embeddings:

ROC-AUC Score: 1.000
DRIFT DETECTED! User queries have changed topic.

12

ROC-AUC Score: 1.000DRIFT DETECTED! User queries have changed topic.

Let’s also try the centroid method one more time:

from sklearn.metrics.pairwise import cosine_distances
import numpy as np


centroid_ref_st = np.mean(X_reference, axis=0).reshape(1, -1)
centroid_prod_st = np.mean(X_production, axis=0).reshape(1, -1)

distance_st = cosine_distances(centroid_ref_st, centroid_prod_st)[0][0]

print(f"Centroid Cosine Distance (SentenceTransformer Embeddings): {distance_st:.4f}")

threshold_st = 0.05  # Adjust threshold as needed
if distance_st > threshold_st:
    print("ALERT: Centroid distance exceeded threshold! System drifting (SentenceTransformer Embeddings).")
else:
    print("System stable: Centroids are aligned (SentenceTransformer Embeddings).")

1234567891011121314151617181920

from sklearn.metrics.pairwise import cosine_distancesimport numpy as np # Centroid Distance for SentenceTransformer embeddings # Calculate centroidscentroid_ref_st = np.mean(X_reference, axis=0).reshape(1, -1)centroid_prod_st = np.mean(X_production, axis=0).reshape(1, -1) # Calculate cosine distancedistance_st = cosine_distances(centroid_ref_st, centroid_prod_st)[0][0] print(f"Centroid Cosine Distance (SentenceTransformer Embeddings): {distance_st:.4f}") # Alerting Logicthreshold_st = 0.05  # Adjust threshold as neededif distance_st > threshold_st:    print("ALERT: Centroid distance exceeded threshold! System drifting (SentenceTransformer Embeddings).")else:    print("System stable: Centroids are aligned (SentenceTransformer Embeddings).")

Output:

Centroid Cosine Distance (SentenceTransformer Embeddings): 0.8719
ALERT: Centroid distance exceeded threshold! System drifting (SentenceTransformer Embeddings).

12

Centroid Cosine Distance (SentenceTransformer Embeddings): 0.8719ALERT: Centroid distance exceeded threshold! System drifting (SentenceTransformer Embeddings).

As we can see, financial/crypto topics and basic IT support can be far apart in the embedding space managed by our chosen model, all-MiniLM-L6-v2, which still yields a high cosine distance — although not nearly as high as in the purely random data scenario.

Wrapping Up #

This article introduced some common techniques used in production machine learning systems to monitor and detect drifts in data represented as vector embeddings. Two of these techniques, namely model-based detection and the centroid distance method, have been illustrated through code examples, aided by Scikit-LLM for embedding generation.