Architecture

Understand Pygrad's internal architecture and design.

Overview

Pygrad is built on a modular architecture that separates concerns into distinct components:

graph TB
    subgraph API["User API Layer"]
        PG[pg.add / pg.search / pg.list]
        CLI[CLI: pygrad]
    end

    subgraph Core["Core Processing"]
        REPO[Repository Manager]
        PARSER[Tree-sitter Parser]
        PROC[Python Processor]
        EXTRACT[Example Extractor]
        XML[XML Generator]
    end

    subgraph Storage["Storage Layer"]
        COGNEE[Cognee Engine]
        VDB[(Vector DB)]
        GDB[(Graph DB)]
    end

    subgraph External["External Services"]
        LLM[LLM Provider]
        EMB[Embedding Model]
    end

    PG --> REPO
    CLI --> REPO
    REPO --> PARSER
    PARSER --> PROC
    PROC --> EXTRACT
    EXTRACT --> XML
    XML --> COGNEE
    COGNEE --> VDB
    COGNEE --> GDB
    COGNEE --> LLM
    COGNEE --> EMB

Key Concepts

Graph RAG

Pygrad uses Graph RAG (Retrieval-Augmented Generation with Graph context):

1. Knowledge Graph: API documentation is stored as a connected graph of entities (classes, methods, functions, examples)
2. Semantic Search: Queries are matched against the graph using vector embeddings
3. Context Extension: Related nodes in the graph are included to provide richer context
4. LLM Generation: An LLM generates the final answer using the retrieved context

Data Flow

GitHub URL → Clone → Parse → Extract → XML → Graph → Search → Answer

Learn More

How It Works

Detailed sequence diagrams showing the complete data flow.

Components

Description of each component and its responsibilities.