A lot of people talk about AI and LLMs. But we see little talk about Embeddings and that’s where a lot of the “AI Magic” works.

At dobror, we use it a lot to improve the way you can consume your emails.

Both at classifying your emails in the correct inbox as well as a Semantic search that we are building.

Understanding Embeddings

Embeddings are a way to represent words, sentences, or even entire documents as vectors of numbers. Think of them as a sophisticated filing system for language. Here’s a simplified example:

• The word “finance” might be represented as [0.2, -0.5, 0.8, …]
• “Receipt” could be [0.3, -0.4, 0.7, …]
• And “invoice” might look like [0.25, -0.45, 0.75, …]

Notice how these vectors are similar? (okay… probably not to easy to identify) But that’s because these words are semantically related. This relationship is what makes embeddings so powerful for understanding language.

The AI will create clusters of related meanings in a 3 dimension matrix. Think about the words: “Elementary”, “Student” and “College”. They all relate to “School” in a way. So the AI will put them near the same cluster in this matrix.

For example:

1. Training: Embeddings are created by training models on large volumes of text data. These models learn to predict words based on their context or vice versa.
2. Dimensionality: While our example used just a few numbers, real-world embeddings often have hundreds or thousands of dimensions to capture nuanced relationships.
3. Similarity Measurement: We can measure the similarity between embeddings using techniques like cosine similarity, allowing us to understand how closely related different pieces of text are.
4. Versatility: Once created, embeddings can be fine-tuned or applied to various tasks, making them incredibly useful across different applications.

Embeddings in Action at dobror.com

At dobror.com, we’ve harnessed the power of embeddings to create a smarter email experience. Here’s how we’re using them:

1. Intelligent Email Classification

We use embeddings to automatically classify incoming emails into relevant categories or “split inboxes.” Here’s how it works:

• When an email arrives, we generate an embedding for its content.
• We compare this embedding to pre-defined category embeddings (e.g., “Finance,” “Personal,” “Work”).
  • You can of course create your own categories or customize them
• The email is then routed to the most similar category.

2. Finance-Specific Classification

One of our most popular features is the automatic classification of finance-related emails:

• We’ve trained our system on a wide range of financial documents, receipts, and emails.
• This allows us to create highly accurate embeddings for financial content.
• When an email resembles a receipt, invoice, or financial statement, it’s automatically routed to your “Finances” split inbox.

3. Personalized Learning

Our system doesn’t just rely on pre-defined categories. It learns from your behavior:

• As you manually move emails between split inboxes, our system updates its understanding.
• The embeddings for each category are fine-tuned based on your specific email patterns.
• Over time, this creates a personalized classification system tailored to your unique needs.

The Benefits of Embedding-Based Classification

By using embeddings at dobror.com, we’ve created several key advantages for our users:

1. Accuracy: Embeddings capture semantic meaning, not just keywords, leading to more accurate classification.
2. Flexibility: The system can understand and categorize new types of emails without explicit programming.
3. Efficiency: Automatic classification saves you time and keeps your inbox organized.
4. Personalization: The system adapts to your specific email patterns and preferences.

Looking to the Future

As we continue to refine our use of embeddings at dobror.com, we’re exploring exciting new possibilities:

• 🔥 Semantic Search: nowadays in most searches in email, you search for “keywords”. So if you want to find emails about your trip to Barcelona, you type: “travel barcelona tickets”.
  But with embeddings, because we can find the meaning of each email, it is possible to have a search by meaning, for example: “my travel with Regina to Spain”. Then the AI will already bring you related emails
  
  🔥Releasing this soon
• Chat with my emails: Using embeddings you can find the related contents and make a “chat with my emails” feature, that you can ask questions about your inbox to the chat.
• Content Summarization: Leveraging embeddings to generate concise summaries of lengthy emails.
• Smart Replies: Suggesting contextually appropriate responses based on email content and your writing style.

Conclusion

By implementing embedding technology at dobror.com, we’ve created an email client that truly understands the content of your messages. Automatically routing your receipts, invoices, and financial statements to a dedicated split inbox is just the beginning of what we can achieve with this powerful tool.

We’re committed to continually improving your email experience, and embeddings are a key part of that mission. As we move forward, we’re excited to bring you even more intelligent features that make email management more intuitive and less time-consuming.

Have you noticed improvements in your dobror.com inbox organization?