Skip to content

knowledge_graph.md

Latest commit

 

History

History
94 lines (70 loc) · 5.12 KB

knowledge_graph.md

File metadata and controls

94 lines (70 loc) · 5.12 KB
Raw

LocalPilot Project Knowledge Graph

LocalPilot is a privacy-first, agent-driven AI pair programmer for Visual Studio. It leverages local LLMs (via Ollama) and deep IDE integration to provide autonomous coding capabilities.

🏗️ System Architecture

LocalPilot follows an Agentic OODA Loop (Observe, Orient, Decide, Act) architecture, powered by a high-performance SQLite storage engine.

graph TD
    UI[LocalPilotChatControl] --> AM[AgentOrchestrator]
    AM --> OS[OllamaService]
    AM --> TR[ToolRegistry]
    AM --> PC[ProjectContextService]
    AM --> RS[RoslynSemanticProvider]
    
    PC -- SQLite RAG --> SS[StorageService]
    AM -- Persistence --> SS
    NEXUS[NexusService] -- Dependency Graph --> SS
    
    RS -- Symbols/Refactoring --> AM
    TR -- File/Shell Access --> AM
    OS -- LLM Inference --> AM
    
    SFIX[SmartFixService] -- Error Detection --> AM
    PERF[PerformanceOptimizer] -- Bottleneck Analysis --> AM
Loading

🧩 Core Services Map

Service Responsibility Key Features
StorageService The Persistence Engine. SQLite WAL Mode, FTS5 Search, Busy-Timeout (5s), Self-Healing Initialization.
AgentOrchestrator The "Brain". Manages the autonomous loop. Performance Shield, Context Budgeting (2k limit), OODA Orientation.
GlobalPriorityGuard Resource Coordinator. Yield-on-Action, 30s Smart Cooldown, CancellationToken-based abortion.
OllamaService LLM Interface. 60s Request Timeout, Circuit Breaker, Native Tool Calling, Text-Only Fallback.
ProjectContextService RAG Layer. Semantic code search. Differential SQLite Sync, Roslyn Chunking, 2MB File Cap.
NexusService Full-Stack Bridge. Maps dependencies. Persistent Graph Storage, Cross-Stack Trace (C# to TS/TSX).
RoslynSemanticProvider Semantic Intelligence. Neighborhood Context, Project-wide Rename, Semantic Diagnostics.
ToolRegistry Capability Layer. File I/O (Safe Overwrite), Grep, Terminal, Unit Testing.
SmartFixService Self-Heal Watchdog. Real-time "Fix with AI" proposals for compilation errors.
PerformanceOptimizer Proactive Optimization. Detects O(n²) loops, blocking async, redundant allocations.

🛠️ Agent Capability Suite (Tools)

The agent has access to several native tools via the ToolRegistry:

  • FileSystem: read_file, write_file, list_directory, delete_file.
  • Search: grep_search (High-perf parallel scan), SearchContextAsync (Hybrid BM25 + Vector Search).
  • Editing: replace_text (Precise block replacement with CRLF/LF normalization).
  • Development: run_terminal (cmd.exe commands), run_tests (auto-detects toolchain).
  • Nexus: trace_dependency (Cross-stack path tracing), analyze_impact (Full-stack change impact analysis).
  • Analysis: list_errors (VS Error List access).

🎨 UI & Design Principles (Ghost UI)

The UI is built using WPF and adheres to the "Ghost UI" design mandate: minimalist, responsive, and natively theme-aware.

  • ChatControl: The primary container. Manages streaming narrative and activity logs.
  • Full UI Virtualization: The entire message container and activity log utilize VirtualizingStackPanel with recycling.
  • Batched Token Pipeline: Fragments are buffered and flushed at 30 FPS via a DispatcherTimer.
  • Incremental Markdown: A structural additive parser (O(1) per frame) ensures smooth text flow during streaming.
  • AgentTurnLayout: Separates the Narrative (LLM text) from the Activity (Tool execution timeline).
  • Human-in-the-Loop (HIL): A security layer requiring user approval for "risky" tools (e.g., file overwrite).
  • Staged Review: Custom diff view for multi-file changes before final acceptance.

⚡ Performance & Resource Management

  1. VRAM Management: Models are automatically unloaded from VRAM after 5-10 minutes of inactivity using keep_alive tuning.
  2. Persistent Caching: Embeddings and FTS5 indices are stored in SQLite, eliminating re-indexing stutters.
  3. Busy-Timeout Protection: SQLite utilizes a 5000ms busy timeout to prevent deadlocks during high-concurrency renames.
  4. Priority Guard: All background services yield CPU/GPU resources immediately when an agent turn starts.
  5. I/O Throttling: RAG indexing is limited to 2 concurrent files to preserve disk bandwidth.
  6. Context Budgeting: Tool outputs are capped at 2,000 characters to preserve LLM reasoning quality.
  7. Solution Visibility: Project Map generation is capped at 1,000 files to ensure O(1) snapshot time.
  8. Layout Debouncing: Scroll operations are throttled to ~25 FPS to prevent layout thrashing.
  9. Self-Healing: The storage engine automatically resets if the database becomes corrupted.

📂 Project Metadata (Internal)

  • App Data Directory: %AppData%\LocalPilot\
  • Persistent Database: localpilot_v2.db (All files, vectors, and graph nodes)
  • Rules: LOCALPILOT.md (Project-specific instructions)
  • Logs: %AppData%\LocalPilot\logs\