Architecture Overview

Architecture Overview

This document provides detailed technical diagrams of Local Deep Research's architecture.

System Architecture

flowchart TB
    subgraph USER["👤 User Interface"]
        WEB[Web Browser<br/>localhost:5000]
    end

    subgraph FLASK["🌐 Flask Backend"]
        API[REST API + WebSocket]
        AUTH[Authentication<br/>CSRF Protection]
        ROUTES[Research Routes]
    end

    subgraph RESEARCH["🔬 Research Engine"]
        STRAT[Strategy Selector<br/>20+ strategies]
        QGEN[Question Generator]
        EXEC[Search Executor]
        SYNTH[Report Synthesizer]
    end

    subgraph LLM["🤖 LLM Providers"]
        direction TB
        LOCAL_LLM[/"🏠 Local LLMs"\]
        OLLAMA[Ollama]
        LMSTUDIO[LM Studio]
        CLOUD_LLM[/"☁️ Cloud LLMs"\]
        OPENAI[OpenAI]
        ANTHROPIC[Anthropic]
        GEMINI[Google Gemini]
        OPENROUTER[OpenRouter<br/>100+ models]
    end

    subgraph SEARCH["🔍 Search Engines"]
        direction TB
        LOCAL_SEARCH[/"🏠 Local Search"\]
        SEARXNG[SearXNG]
        ELASTIC[Elasticsearch]
        LIBRARY[Document Library]
        WEB_SEARCH[/"🌐 Web Search"\]
        TAVILY[Tavily]
        BRAVE[Brave]
        DDG[DuckDuckGo]
        ACADEMIC[/"📚 Academic"\]
        ARXIV[ArXiv]
        PUBMED[PubMed]
        SEMANTIC[Semantic Scholar]
    end

    subgraph STORAGE["💾 Storage Layer"]
        SQLCIPHER[(SQLCipher DB<br/>AES-256 Encrypted)]
        VECTORS[(Vector Store<br/>Embeddings)]
        FILES[(File Storage<br/>PDFs & Docs)]
    end

    subgraph OUTPUT["📄 Output"]
        MD[Markdown]
        PDF[PDF Export]
        LATEX[LaTeX]
        QUARTO[Quarto]
        RIS[RIS/BibTeX]
    end

    WEB <--> API
    API --> AUTH
    AUTH --> ROUTES
    ROUTES --> STRAT

    STRAT --> QGEN
    QGEN --> EXEC
    EXEC --> SYNTH

    EXEC <--> OLLAMA & LMSTUDIO
    EXEC <--> OPENAI & ANTHROPIC & GEMINI & OPENROUTER

    EXEC <--> SEARXNG & ELASTIC & LIBRARY
    EXEC <--> TAVILY & BRAVE & DDG
    EXEC <--> ARXIV & PUBMED & SEMANTIC

    SYNTH --> SQLCIPHER
    LIBRARY <--> VECTORS
    LIBRARY <--> FILES

    SYNTH --> MD & PDF & LATEX & QUARTO & RIS

    LOCAL_LLM ~~~ OLLAMA
    CLOUD_LLM ~~~ OPENAI
    LOCAL_SEARCH ~~~ SEARXNG
    WEB_SEARCH ~~~ TAVILY
    ACADEMIC ~~~ ARXIV

    style USER fill:#e1f5fe
    style FLASK fill:#fff3e0
    style RESEARCH fill:#f3e5f5
    style LLM fill:#e8f5e9
    style SEARCH fill:#fce4ec
    style STORAGE fill:#fff8e1
    style OUTPUT fill:#e0f2f1

Research Flow

flowchart LR
    subgraph INPUT["1️⃣ Input"]
        Q[Research Query]
    end

    subgraph ITERATE["2️⃣ Iterative Research"]
        direction TB
        GEN[Generate<br/>Questions]
        SEARCH[Parallel<br/>Search]
        ANALYZE[Analyze<br/>Results]
        GEN --> SEARCH --> ANALYZE
        ANALYZE -.->|"Need more info?"| GEN
    end

    subgraph SYNTHESIZE["3️⃣ Synthesis"]
        REPORT[Generate<br/>Report]
        CITE[Add<br/>Citations]
        REPORT --> CITE
    end

    subgraph OUTPUT["4️⃣ Output"]
        RESULT[Final Report<br/>with Sources]
    end

    Q --> GEN
    ANALYZE --> REPORT
    CITE --> RESULT

    style INPUT fill:#e3f2fd
    style ITERATE fill:#f3e5f5
    style SYNTHESIZE fill:#e8f5e9
    style OUTPUT fill:#fff3e0

Deployment Options

flowchart TB
    subgraph FULL_LOCAL["🏠 Fully Local (Maximum Privacy)"]
        direction LR
        L_LDR[LDR] <--> L_OLLAMA[Ollama]
        L_LDR <--> L_SEARX[SearXNG]
        L_LDR <--> L_DB[(Encrypted DB)]
    end

    subgraph HYBRID["⚡ Hybrid (Balanced)"]
        direction LR
        H_LDR[LDR] <--> H_OLLAMA[Ollama]
        H_LDR <-->|"Web Search"| H_CLOUD[Cloud APIs]
        H_LDR <--> H_DB[(Encrypted DB)]
    end

    subgraph CLOUD["☁️ Cloud-Powered (Maximum Speed)"]
        direction LR
        C_LDR[LDR] <--> C_OPENAI[OpenAI/Claude]
        C_LDR <--> C_TAVILY[Tavily/Brave]
        C_LDR <--> C_DB[(Encrypted DB)]
    end

    style FULL_LOCAL fill:#e8f5e9
    style HYBRID fill:#fff3e0
    style CLOUD fill:#e3f2fd

Feature Map

mindmap
  root((Local Deep<br/>Research))
    Research
      Quick Summary
      Detailed Reports
      Follow-up Questions
      20+ Strategies
    Search Sources
      Web Search
        SearXNG
        Tavily
        Brave
        DuckDuckGo
      Academic
        ArXiv
        PubMed
        Semantic Scholar
      Local
        Document Library
        Collections
        Elasticsearch
    LLM Support
      Local
        Ollama
        LM Studio
      Cloud
        OpenAI
        Anthropic
        Google
        OpenRouter
    Output
      Markdown
      PDF
      LaTeX
      Quarto
      RIS/BibTeX
    Features
      News Subscriptions
      Cost Analytics
      Benchmarking
      RAG Search
      Per-User Encryption

Component Details

LLM Providers

Search Engines

Output Formats

Knowledge Loop: Research → Library → Future Research

One of LDR's powerful features is the ability to build a personal knowledge base that improves future research.

flowchart TB
    subgraph RESEARCH["1️⃣ Research"]
        Q[Your Question] --> Engine[Research Engine]
        Engine --> Results[Results + Sources]
    end

    subgraph DOWNLOAD["2️⃣ Download"]
        Results --> Track[Track Sources]
        Track --> Queue[Download Queue]
        Queue --> Extract[Download & Extract Text]
    end

    subgraph LIBRARY["3️⃣ Library"]
        Extract --> Store[(Document Storage)]
        Store --> Organize[Organize into Collections]
    end

    subgraph INDEX["4️⃣ Index"]
        Organize --> Chunk[Chunk Documents]
        Chunk --> Embed[Generate Embeddings]
        Embed --> FAISS[(Vector Index)]
    end

    subgraph REUSE["5️⃣ Reuse"]
        FAISS --> SearchEngine[Collection as Search Engine]
        SearchEngine --> Q
    end

    style RESEARCH fill:#e3f2fd
    style DOWNLOAD fill:#fff3e0
    style LIBRARY fill:#e8f5e9
    style INDEX fill:#f3e5f5
    style REUSE fill:#fce4ec

How It Works

1. Research Completes → Sources are tracked in ResearchResource table
2. Download Sources → Click "Get All Research PDFs" to queue downloads
   • Smart downloaders for ArXiv, PubMed, Semantic Scholar, etc.
   • Automatic text extraction from PDFs
3. Build Library → Documents stored in encrypted database
   • Deduplication via content hash
   • Multiple storage modes: database (encrypted), filesystem, text-only
4. Create Collections → Organize documents by topic/project
   • Each collection can have different embedding settings
   • Documents can belong to multiple collections
5. Index for Search → Generate vector embeddings
   • Configurable chunk size and overlap
   • FAISS index for fast similarity search
6. Use in Future Research → Select collection as search engine
   • RAG search finds relevant passages
   • Results cite back to your documents

Key Components

Storage Options

Technical Analysis & Project Health

Last updated: December 2024

This section provides a comprehensive technical analysis of the codebase, including quality metrics, architecture patterns, and project health indicators.

Project Statistics

Architecture Patterns

Extensibility Design

The codebase follows a consistent pattern for extensibility:

┌─────────────────────────────────────────────────────────────┐
│                    Abstract Base Classes                     │
├─────────────────────────────────────────────────────────────┤
│ BaseSearchEngine      │ Common interface for 25+ engines    │
│ BaseSearchStrategy    │ Strategy pattern for research       │
│ BaseCitationHandler   │ Citation processing abstraction     │
│ BaseQuestionGenerator │ Question generation interface       │
│ BaseExporter          │ Export format abstraction           │
└─────────────────────────────────────────────────────────────┘

Search Engine Plugin System

New search engines can be added by:

1. Creating a class inheriting from BaseSearchEngine
2. Placing it in web_search_engines/engines/
3. Auto-discovery handles registration

# Example: Adding a new search engine
class SearchEngineCustom(BaseSearchEngine):
    def run(self, query: str) -> List[Dict]:
        # Implementation
        pass

LLM Provider Integration

Supports 9 LLM providers with auto-discovery:

• Ollama (local)
• LM Studio (local)
• OpenAI
• Anthropic
• Google Gemini
• OpenRouter (100+ models)
• DeepSeek
• Mistral
• Groq

Quality Ratings

pie title Project Health Ratings
    "Code Quality (86)" : 86
    "Security (93)" : 93
    "Extensibility (93)" : 93
    "Performance (85)" : 85
    "UX/DX (83)" : 83
    "Maintenance (89)" : 89

Detailed Ratings

Security Architecture

flowchart TB
    subgraph SECURITY["🔐 Security Layers"]
        direction TB
        AUTH[Flask Session Auth<br/>CSRF Protection]
        ENCRYPT[SQLCipher<br/>AES-256 Encryption]
        SCAN[22+ Security Scanners<br/>CodeQL, Semgrep, Bandit...]
        PIN[Full SHA Pinning<br/>All 57 CI Workflows]
    end

    AUTH --> ENCRYPT
    ENCRYPT --> SCAN
    SCAN --> PIN

    style SECURITY fill:#e8f5e9

Security Features:

• Per-user encrypted databases (SQLCipher with AES-256)
• Full GitHub Action SHA pinning (not tag-based)
• Comprehensive CI security scanning:
  • CodeQL (Python, JavaScript)
  • Semgrep (custom rulesets)
  • Bandit (Python security)
  • Trivy (container scanning)
  • Dependency review
  • Secret scanning

Performance Optimizations

Areas for Improvement

While the project scores highly overall, these areas have room for growth:

1. Integration Testing - More end-to-end tests for full research workflows
2. API Documentation - OpenAPI/Swagger spec for REST endpoints
3. Metrics Dashboard - Prometheus/Grafana integration for monitoring
4. Container Optimization - Multi-stage Docker builds for smaller images
5. Async Architecture - Migration to async/await for I/O-bound operations

Key Source Files

Contributing to Architecture

When extending the system:

1. Adding Search Engines: Inherit from BaseSearchEngine, implement run() method
2. Adding Strategies: Inherit from BaseSearchStrategy, implement analyze_topic() method
3. Adding LLM Providers: Add to config/llm_config.py with proper initialization
4. Adding Export Formats: Inherit from base exporter pattern in utilities/

See CONTRIBUTING.md for detailed guidelines.