Skip to content

architecture.md

Latest commit

 

History

History
43 lines (34 loc) · 2.01 KB

architecture.md

File metadata and controls

43 lines (34 loc) · 2.01 KB

Mimir Architecture

Mimir is a code intelligence engine designed to index source code into a high-fidelity knowledge graph and expose it to AI agents via the Model Context Protocol (MCP).

High-Level Overview

graph TD
    Repo[Source Code] --> Walker[Parallel Walker]
    Walker --> Parser[Tree-sitter Workers]
    Parser --> SymbolTable[Global Symbol Table]
    SymbolTable --> Resolver[Scope-aware Resolver]
    Resolver --> Store[SQLite Knowledge Graph]
    Store --> Embedder[Async Embedder]
    Store --> Analysis[Louvain Clustering & BFS Tracing]
    Store --> MCP[MCP Server & HTTP Bridge]
    MCP --> Agent[AI Agent]
Loading

Core Components

1. Parallel Walker & Parser

The internal/walker package performs concurrent file enumeration while respecting .gitignore. The internal/parser package uses a pre-allocated pool of tree-sitter workers to extract symbols (Functions, Classes, Methods, etc.) without the overhead of re-creating parsers for every file.

2. Scope-aware Resolver

The internal/resolver performs a two-pass algorithm:

  • Pass 1: Build a global Symbol Table (Trie-based) of all exported and internal symbols.
  • Pass 2: Resolve call sites and references to their definitions across file boundaries, assigning confidence scores based on resolution certainty.

3. Knowledge Graph Store

Mimir uses SQLite (modernc.org/sqlite) with sqlite-vec for high-performance storage.

  • Nodes: Symbols with metadata (file, line, exported status).
  • Edges: Relationships (CALLS, IMPORTS, EXTENDS, IMPLEMENTS).
  • BM25: Full-text search index.
  • Vector: HNSW-based semantic search.

4. Analysis Engines

  • Community Detection: Uses the Louvain algorithm to group related code into "clusters".
  • Process Tracing: Uses BFS to identify end-to-end execution flows (processes) starting from entry points.

5. Transport Layer

  • MCP Server: Implements the Model Context Protocol over stdio.
  • HTTP Bridge: Exposes the knowledge graph via a REST API for remote integration.