Skip to main content

AI Chat Tab

The AI Chat tab provides an interactive natural-language interface to your heap analysis. Ask questions about the heap dump and receive answers grounded in the actual analysis data.

What You See​

A chat interface with a message input at the bottom and a scrollable message history above it:

┌──────────────────────────────────────────────────────┐
│  🤖 Based on the heap analysis, the primary memory   │
│     consumer is com.example.cache.DataCache at        │
│     512 MB (31.4% of heap). This single HashMap-      │
│     backed cache retains nearly a third of all         │
│     reachable memory...                                │
│                                                        │
│  👤 What are the top 3 things I should fix first?     │
│                                                        │
│  🤖 Here are the top 3 recommendations:               │
│     1. **DataCache** (512 MB) — Add LRU eviction...   │
│     2. **Duplicate strings** (35 MB) — Enable JVM...  │
│     3. **Empty collections** (9.6 MB) — Lazy init...  │
├──────────────────────────────────────────────────────┤
│  [Type your question here...]            [Send]       │
└──────────────────────────────────────────────────────┘

How It Works​

Context Injection​

On the first message, HeapLens prepends a structured summary of the heap analysis as context for the LLM. This includes:

  • Heap Summary — Total size, object counts, GC root count
  • Top 15 Objects — Class name, type, shallow size, retained size
  • All Leak Suspects — Severity, class, percentage, description, dependency info
  • Top 20 Classes — From the class histogram
  • Waste Analysis — Total waste, top 5 duplicate strings

This context is formatted as markdown tables (~2-3K tokens) and included as a system message, so the LLM can answer questions accurately about your specific heap dump.

Streaming Responses​

Responses are streamed token-by-token for fast feedback — you see the answer being written in real time rather than waiting for the complete response.

Conversation History​

The chat maintains a per-editor conversation history (up to 20 exchanges). Follow-up questions reference the context of previous answers.

Configuration​

The AI Chat requires an API key. Configure it in VS Code settings:

SettingDescriptionExample
heaplens.llm.providerLLM provider"anthropic" or "openai"
heaplens.llm.apiKeyAPI key"sk-ant-..."
heaplens.llm.baseUrlCustom API endpoint (optional)"https://my-proxy.example.com"
heaplens.llm.modelModel override (optional)"claude-sonnet-4-20250514"

Without an API key, the chat tab shows a configuration prompt.

Example Questions​

Here are effective questions to ask about your heap dump:

Diagnostic Questions​

  • "What is the biggest memory consumer and why?"
  • "Is there a memory leak? What evidence supports that?"
  • "Why is byte[] the largest class by shallow size?"
  • "Are there any suspicious retention patterns?"

Actionable Questions​

  • "What are the top 3 things I should fix to reduce memory usage?"
  • "How can I reduce the 35 MB of duplicate strings?"
  • "What code change would have the most impact on heap size?"
  • "Should I add a cache eviction policy? What size limit would you recommend?"

Investigative Questions​

  • "Explain the relationship between DataCache and the HashMap retaining 510 MB"
  • "Are the 45,000 UserSession objects expected for a service with 100 concurrent users?"
  • "What are the empty collections and are they worth fixing?"
  • "Compare the leak suspects — which is the root cause vs. a symptom?"

Copilot Chat Integration​

HeapLens also registers as a VS Code Chat Participant. If you have GitHub Copilot Chat installed, you can use the @heaplens mention in the Copilot Chat panel:

@heaplens What's causing the high memory usage?

This uses the same analysis context injection but integrates with the Copilot Chat UI rather than HeapLens's built-in chat tab.

Limitations​

  • The LLM sees a summary of the analysis data, not the raw heap dump. It cannot access individual objects that are not in the top 15 or the histogram.
  • For deep investigation (e.g., "what is object #1234567 retaining?"), use the Dominator Tree tab directly.
  • Responses are generated by the LLM and may occasionally be inaccurate. Always verify recommendations against the actual data in other tabs.