question answering of codebases - chunhualiao/public-docs GitHub Wiki

Question: what are best open source solutions to answer questions about a large scale C++ codebase, possibly using Large language models and RAG techniques? What are the standardized benchmarks and leaderboards for this codebase question answering task?

survey the field

DeepWiki-Open

I'll create a diagram to explain how the question answering system works in DeepWiki's codebase .

graph TD
    subgraph "User Interface"
        A["User asks question about codebase"]
        B["Frontend sends query via WebSocket/HTTP"]
    end
    
    subgraph "API Layer"
        C["WebSocket Handler<br/>(/ws/chat)"]
        D["HTTP Stream Handler<br/>(/chat/completions/stream)"]
    end
    
    subgraph "RAG Processing Pipeline"
        E["Create RAG Instance"]
        F["Prepare Repository Retriever"]
        G["Query Vector Database<br/>(FAISS embeddings)"]
        H["Retrieve Relevant Code Snippets"]
        I["Group Documents by File Path"]
        J["Format Context with Headers"]
    end
    
    subgraph "Prompt Construction"
        K["Build Comprehensive Prompt"]
        L["Include Repository Context"]
        M["Add Query + Conversation History"]
        N["Apply Language Preferences"]
    end
    
    subgraph "AI Model Providers"
        O["Google Gemini"]
        P["OpenAI"]
        Q["OpenRouter"]
        R["Ollama Local"]
        S["Azure OpenAI"]
    end
    
    subgraph "Response Processing"
        T["Stream AI Response"]
        U["Handle Token Limits"]
        V["Apply Deep Research Logic"]
        W["Return Formatted Answer"]
    end
    
    subgraph "Storage Layer"
        X["Code Embeddings<br/>(~/.adalflow/databases/)"]
        Y["Repository Cache<br/>(~/.adalflow/repos/)"]
        Z["Conversation Memory"]
    end
    
    A --> B
    B --> C
    B --> D
    
    C --> E
    D --> E
    
    E --> F
    F --> G
    G --> X
    G --> H
    H --> I
    I --> J
    
    J --> K
    K --> L
    L --> M
    M --> N
    
    N --> O
    N --> P
    N --> Q
    N --> R
    N --> S
    
    O --> T
    P --> T
    Q --> T
    R --> T
    S --> T
    
    T --> U
    U --> V
    V --> W
    
    F --> Y
    K --> Z
    
    W --> B
    B --> A
    
    classDef userLayer fill:#e1f5fe
    classDef apiLayer fill:#f3e5f5
    classDef ragLayer fill:#e8f5e8
    classDef promptLayer fill:#fff3e0
    classDef aiLayer fill:#fce4ec
    classDef responseLayer fill:#f1f8e9
    classDef storageLayer fill:#f5f5f5
    
    class A,B userLayer
    class C,D apiLayer
    class E,F,G,H,I,J ragLayer
    class K,L,M,N promptLayer
    class O,P,Q,R,S aiLayer
    class T,U,V,W responseLayer
    class X,Y,Z storageLayer

Key Components Explained

RAG Processing Pipeline: The system creates code embeddings stored in ~/.adalflow/databases/ and uses FAISS for vector similarity search to find relevant code snippets 1 .

Dual API Architecture: Questions can be processed through either WebSocket connections (/ws/chat) or HTTP streaming (/chat/completions/stream) endpoints, both implementing the same core RAG logic .

Multi-Provider Support: The system supports multiple AI providers including Google Gemini, OpenAI, OpenRouter, Azure OpenAI, and local Ollama models 2 .

Context-Aware Responses: Retrieved code snippets are grouped by file path and formatted with clear headers before being sent to the AI model for context-aware documentation generation 3 .

Advanced Features: The system includes Deep Research capabilities for complex queries and file-specific query handling when users specify particular file paths .

Notes

This diagram illustrates the complete flow from user question to AI-generated answer, showing how DeepWiki combines vector search with large language models to provide accurate, context-aware responses about codebases. The system maintains conversation history and supports multiple languages based on user preferences.

Wiki pages you might want to explore: