Core Abstractions

eLLMental uses different 3rd party components and APIs and provides a unified interface. To ensure extensibility and avoid tight coupling with any specific API, the library provides a series of abstract classes that define the expected interface for these components to work with eLLMental. To use eLLMental, you can provide your own implementation or use one of the built-in concrete implementations.

Embedding object

In eLLMental embeddings are represented by the Embedding record, which has the following attributes:

• id: An unique identifier of the embedding.
• vector: A numeric vector that represents the semantic location of the text.
• metadata: Additional information associated with the embedding. It can be used to store the original text, the model used to generate the embedding, or any other information you may find useful.

public record Embedding(
        UUID id,
        List<Double> vector,
        Map<String, String> metadata
) {}

EmbeddingsGenerationModel

This abstract class defines the interface expected by eLLMental for a valid embeddings generation model.

public abstract class EmbeddingsGenerationModel {
    public abstract Embedding generateEmbedding(String text);
}

OpenAIEmbeddingsGenerationModel

eLLMental provides an implementation to use OpenAI's embeddings model. This model is only accessible via API, so you'll need to initialize it with a valid OpenAI API key.

EmbeddingsGenerationModel openAIModel = new OpenAIEmbeddingsGenerationModel("YOUR_OPENAI_API_KEY");

// You'll rarely need to interact directly with the `openAIModel`, but you can use it to generate an embedding object:
Embedding embedding = openAIModel.generateEmbedding("Sample string");

The OpenAI embeddings generator will automatically include the original text, the timestamp and the model used to generate the embedding in the metadata.

EmbeddingsStore

This abstract class defines the expected interface for a persistence mechanism capable of storing and querying embeddings:

public abstract class EmbeddingsStore {
    public abstract void store(Embedding embedding);
    public abstract List<Embedding> similaritySearch(Embedding reference, int limit);
}

PineconeEmbeddingsStore

eLLMental provides a concrete implementation for Pinecone, which requires defining a URL, an API Key, and a namespace.

EmbeddingsStore pineconeStore = new PineconeEmbeddingsStore("YOUR_PINECONE_URL", "YOUR_PINECONE_API_KEY", "YOUR_PINECONE_NAMESPACE");

// You can now insert, fetch, delete or perform similarity searches using the pineconeStore instance:
pineconeStore.store(someEmbedding);
List<Embedding> similarEmbeddings = pineconeStore.similaritySearch(referenceEmbedding, 5);

// Get back an embedding by id without a namespace
UUID someEmbeddingId = someEmbedding.id()
Embedding embedding = pineconeStore.get(someEmbeddingId);

// Get a namespaced embedding by id
UUID anotherEmbeddingId = UUID.fromString("01870603-f211-7b9a-a7ea-4a98f5320ff8")
Embedding anotherEmbedding = pineconeStore.get(anotherEmbeddingId, "my-namespace");

// Delete an embedding by id
pineconeStore.delete(anotherEmbeddingId);