AI编码Agent代码知识图谱预索引方案设计器

You are an expert in code intelligence and knowledge graph systems. I need you to design a Pre-indexed Code Knowledge Graph system that allows AI coding agents (Claude Code, Codex, Cursor) to understand a codebase BEFORE starting a conversation, drastically reducing token usage and tool calls.

Project Context

• Language(s): [e.g., TypeScript + Python]
• Repo size: [e.g., 500 files / 100K LOC]
• Framework(s): [e.g., Next.js + FastAPI]
• Current pain: [e.g., Agent reads 50+ files per task, burning 200K tokens]

Please Design:

1. Graph Schema

Define node types and relationships:

• Files, Functions, Classes, Interfaces, Modules
• Import/Export edges, Call edges, Inheritance edges
• Semantic clusters (feature domains, layers)

2. Indexing Pipeline

• AST parsing strategy per language
• Symbol resolution and cross-file reference tracking
• Incremental update on git diff (only re-index changed files)
• Embedding generation for semantic search
• Storage format (JSON-LD, SQLite, Neo4j, or custom)

3. Query Interface for Agents

• Natural language → graph traversal translation
• "Find all callers of function X" in O(1)
• "What files are affected if I change interface Y?"
• "Show me the data flow from API endpoint to database"
• Context window budget-aware result truncation

4. Agent Integration

• How to inject graph context into system prompt
• Token budget allocation: graph summary vs raw code
• Lazy loading strategy (summary first, details on demand)
• Cache invalidation on file changes

5. Metrics & Evaluation

• Token savings percentage vs naive file reading
• Tool call reduction ratio
• Answer accuracy impact (does less context hurt quality?)
• Index build time and storage overhead

6. Implementation Roadmap

Provide a 3-phase plan:

• Phase 1: MVP with static analysis only
• Phase 2: Add semantic embeddings
• Phase 3: Real-time incremental updates

Output as a technical design document with architecture diagrams (Mermaid), example queries, and concrete token savings estimates.