Why You Need This

You're using Copilot, ChatGPT, Cursor, or Claude — and they're great at generating code. But when you need to verify, interrogate, or deep-dive into their responses, you're stuck asking follow-up questions in the same chat window.

That's a problem.

The Context Drift Trap

When you ask follow-up questions inside your main AI chat:

1. Context gets polluted — Your 2,000-word architectural response gets buried under new questions
2. AI starts hallucinating — The agent confuses follow-up questions with code editing instructions
3. You burn tokens — Re-sending entire transcripts for simple clarifications is expensive
4. You can't verify — Is the AI's follow-up explanation accurate, or did it just make something up?

The Solution: Scoped Q&A Sandbox

QA Assistant gives you an isolated console to interrogate any AI response — without touching your main chat.

How It Works

┌─────────────────────────────────────────────────────────────┐
│  1. CAPTURE                                                  │
│  Copy any AI response → Press Ctrl+Alt+Q                    │
│  (or right-click → "Ask About Selection")                   │
└─────────────────────────────────────────────────────────────┘
                            ↓
┌─────────────────────────────────────────────────────────────┐
│  2. INTERROGATE                                              │
│  Ask follow-up questions in the sidebar console             │
│  "Explain this simpler"  •  "What are the trade-offs?"      │
│  "Show me an example"    •  "Is this safe to use?"          │
└─────────────────────────────────────────────────────────────┘
                            ↓
┌─────────────────────────────────────────────────────────────┐
│  3. GET GROUNDED ANSWERS                                     │
│  AI retrieves context from YOUR captured text               │
│  Citations included • No hallucination • 100% local         │
└─────────────────────────────────────────────────────────────┘

Key Features

Local & Private

Your code never leaves your machine. Embeddings run 100% locally via ONNX — no cloud vector databases, no telemetry, no tracking.

Works With Any AI

Copilot, ChatGPT, Cursor, Claude, DeepSeek, local Ollama — capture responses from any AI tool and interrogate them here.

Smart Retrieval

Hybrid search combines semantic understanding (vector similarity) with keyword matching (BM25) for accurate context retrieval.

Real-Time Streaming

Watch answers generate token-by-token with live streaming — no waiting for batch responses.

Multi-Turn Context

Ask follow-up questions naturally. The system remembers your last 3-4 questions for conversational flow.

Quick Start

Install

Search "QA Assistant" in the VS Code Extensions panel (Ctrl+Shift+X) and click Install.

Setup (2 minutes)

1. Configure provider — Ctrl+Shift+P → QA Assistant: Configure LLM Provider
2. Set API key — Ctrl+Shift+P → QA Assistant: Set LLM API Key
   • No key needed for local Ollama
3. Capture & ask — Copy any AI response, press Ctrl+Alt+Q, ask a question

Supported Providers

• OpenRouter (DeepSeek, Llama, Claude, Mistral — 400+ models)
• DeepSeek API directly
• Ollama / LM Studio / vLLM (local)
• OpenAI / Azure OpenAI (via custom endpoint)

Commands

Privacy & Security

• Zero Telemetry — No analytics, no tracking, no metrics
• Local Embeddings — All vectorization happens on your device
• Encrypted Keys — API keys stored in OS keychain via VS Code SecretStorage
• No Code Transmission — Your code never leaves your machine for indexing

Documentation

• CHANGELOG.md — Release notes
• LICENSE — MIT License

Questions or Issues?

Open an issue on GitHub.