Skip to content

FAQ.md

Latest commit

 

History

History
104 lines (68 loc) · 3.82 KB

FAQ.md

File metadata and controls

104 lines (68 loc) · 3.82 KB
Raw

V.I.S.O.R. — Frequently Asked Questions

What is V.I.S.O.R.?

V.I.S.O.R. (Visual Intelligence System for Orchestrated Reasoning) is a local-first MCP server that gives AI coding agents precise, ranked codebase context instead of letting them blindly search through files.

How is this different from just using embeddings?

Embeddings only measure semantic similarity — "does this code look like it's about authentication?" V.I.S.O.R. combines 5 signals:

  1. Embedding similarity — semantic relevance (like pure embedding search)
  2. Exact symbol match — direct name matches in function/class names
  3. Co-location — proximity to the top semantic hit
  4. Dependency graph — architectural connections between files
  5. Recency — recently modified files are more likely relevant to active bugs

Pure embedding search gives you semantically similar code. V.I.S.O.R. gives you architecturally relevant code.

Does V.I.S.O.R. replace my AI agent?

No. V.I.S.O.R. is a co-processor for your AI agent. It doesn't generate code — it helps your existing AI (Antigravity, Claude, GPT-4, Gemini) make better decisions by giving it the right context. Think of it as giving your AI a map instead of making it wander.

What IDEs are supported?

IDE Support Level Method
Google Antigravity ✅ Full (MCP + HUD) MCP config + VSIX extension
VS Code ✅ Full (MCP + HUD) MCP config + VSIX extension
Cursor ✅ MCP tools MCP config
Claude Code ✅ MCP tools MCP config

What languages can V.I.S.O.R. index?

Currently: Python, TypeScript, JavaScript, TSX, Go, Rust, Java, C, and C++ (9 languages).

V.I.S.O.R. uses Tree-sitter for AST parsing, so adding new languages is straightforward (~15 lines of code) — see CONTRIBUTING.md.

Is my code sent to any server?

No. V.I.S.O.R. is entirely local-first:

  • Code is indexed into a SQLite database on your machine
  • Embeddings are generated locally using all-MiniLM-L6-v2
  • Communication with your IDE happens over stdio (no network ports)
  • No data ever leaves your machine

How much does it reduce token usage?

In typical codebases, V.I.S.O.R. achieves 70-85% token reduction compared to naive full-file reads. The exact number depends on:

  • Codebase size
  • Query specificity
  • Active skill strategy

Every response includes exact metrics:

{
  "estimated_tokens_without": 11400,
  "estimated_tokens_with": 2180,
  "reduction_percent": 80.9
}

What are Skills?

Skills are pre-loaded strategies that change how V.I.S.O.R. scores and retrieves context. Each skill can:

  • Override the intent classifier (e.g., force BUG_FIX mode)
  • Adjust individual scoring weights (e.g., boost dependency proximity)
  • Suggest specific follow-up tools

V.I.S.O.R. ships with 4 built-in skills: bug-fixer, architecture-explainer, refactor-assistant, and performance-optimizer. You can create your own via the add_custom_skill MCP tool.

Do I need the VS Code extension, or just the MCP server?

The MCP server is the core product — it provides all 16 tools to your AI agent. The VS Code extension is optional and adds:

  • 3D WebGPU visualization of your codebase architecture
  • Real-time telemetry (Agent Context Burn, Graph Scale, Drift Alerts)
  • Skill management UI

You can use V.I.S.O.R. purely via MCP config + CLI without the extension.

How fast is indexing?

  • First index: 5-15 seconds for a medium codebase (~500 files)
  • Subsequent changes: Incremental — only re-indexes files with changed SHA-256 hashes
  • Model download: First boot downloads all-MiniLM-L6-v2 (~80MB), cached locally after that