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
# Simulating 384-dimensional embeddings (e.g. standard sentence-transformers output) n_samples = 500 n_features = 384
# 1. Referencing Embeddings (Baseline / Training Data) # Imagine this is the data your LLM/Vector DB was originally populated with np.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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 |
from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import train_test_split from sklearn.metrics import roc_auc_score
# 1. Assigning labels: 0 for reference, baseline embeddings; 1 for production embeddings y_reference = np.zeros(n_samples) y_production = np.ones(n_samples)
# 2. Combining into a single dataset X_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 detector X_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 classifier drift_classifier = RandomForestClassifier(n_estimators=50, max_depth=5, random_state=42) drift_classifier.fit(X_train, y_train)
# 5. Evaluating the classifier using ROC-AUC 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}”)
# 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. |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 |
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 vector centroid_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 apart distance = 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 variance threshold = 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. |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 |
from sentence_transformers import SentenceTransformer from google.colab import userdata from skllm.config import SKLLMConfig
# Securely extract the Groq API Key you may have previously stored in Colab secrets groq_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 model vectorizer = SentenceTransformer(‘all-MiniLM-L6-v2’)
# Baseline raw texts and production texts, clearly with a drastic topic shift 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
# Converting text to embeddings X_reference = vectorizer.encode(texts_reference) X_production = vectorizer.encode(texts_production)
# Implementing Embedding Drift Detection Logic # Assign labels: 0 for reference, 1 for production y_reference = np.zeros(len(X_reference)) y_production = np.ones(len(X_production))
# Combining datasets X_combined = np.vstack((X_reference, X_production)) y_combined = np.hstack((y_reference, y_production))
# Training the domain classifier 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)
# Calculating drift using ROC-AUC 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.”) |
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. |
Let’s also try the centroid method one more time:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 |
from sklearn.metrics.pairwise import cosine_distances import numpy as np
# Centroid Distance for SentenceTransformer embeddings
# Calculate centroids centroid_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 distance distance_st = cosine_distances(centroid_ref_st, centroid_prod_st)[0][0]
print(f“Centroid Cosine Distance (SentenceTransformer Embeddings): {distance_st:.4f}”)
# Alerting Logic 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).”) |
Output:
|
Centroid Cosine Distance (SentenceTransformer Embeddings): 0.8719 ALERT: 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.