Aether SDK Architecture

Aether SDK is a modular, protocol-driven framework for building intelligent agents with persistent memory and knowledge graph capabilities.

High-Level Architecture

The system is designed with a layered approach, ensuring that core logic is decoupled from specific provider implementations.

graph TD
    subgraph Orchestration
        B[BrainResource]
        M[AgentMemory]
    end

    subgraph "Functional Modules"
        I[IngestionResource]
        R[RetrievalResource]
        P[ParsingResource]
        E[ExtractionResource]
        D[DistillationResource]
    end

    subgraph "Core Interfaces"
        LP[LLMAdapterProtocol]
        EP[EmbeddingAdapterProtocol]
        VP[VectorStoreProtocol]
        GP[GraphStoreProtocol]
    end

    subgraph "Integrations"
        OA[OpenAI]
        LL["LiteLLM / OpenRouter"]
        QD[Qdrant]
        CH[Chroma]
        NJ[Neo4j]
    end

    Orchestration --> "Functional Modules"
    "Functional Modules" --> "Core Interfaces"
    "Core Interfaces" --> "Integrations"

System Layers

1. Orchestration Layer

• BrainResource: The high-level interface for Question Answering. It orchestrates context retrieval and answer distillation.
• AgentMemory: Manages different memory types—Working (volatile), Episodic (semantic vector-based), and Semantic (graph-based).

2. Resource Layer

Functional modules that handle specific tasks: - Ingestion: Manages the pipeline of parsing, chunking, and storing documents. - Retrieval: Performs hybrid search across vector and graph stores. - Parsing: Converts various file types into clean text. - Extraction: Uses LLMs to extract entities and relations for the Knowledge Graph. - Distillation: Refines retrieved context into a final answer.

3. Protocol Layer

Pure interface definitions that allow for plug-and-play integrations. Every component interacts with external services (LLMs, Databases) only via these protocols.

4. Adapter Layer

Concrete implementations of protocols for specific services (e.g., OpenAIAdapter, QdrantAdapter).

Data Flow

Ingestion Flow

1. Document is received.
2. Parser extracts text.
3. Chunker splits text into manageable pieces.
4. Embedder generates vectors for chunks.
5. Vector Store saves vectors and metadata.
6. Extractor pulls entities/relations (optional).
7. Graph Store saves the knowledge structure (optional).

Query Flow

1. User Query is received.
2. Embedder vectorizes the query.
3. Retriever performs a vector search + optional graph traversal.
4. Distiller (LLM) generates an answer using the retrieved context.