wiki/content/pages/docs-ask-ai.md

title, slug, published, order, showInNav, layout, rightSidebar, showImageAtTop, authorName, authorImage, image, showFooter, docsSection, docsSectionOrder, docsSectionGroup, docsSectionGroupIcon

title	slug	published	order	showInNav	layout	rightSidebar	showImageAtTop	authorName	authorImage	image	showFooter	docsSection	docsSectionOrder	docsSectionGroup	docsSectionGroupIcon
Ask AI	docs-ask-ai	true	2	false	sidebar	true	true	Markdown	/images/authors/markdown.png	/images/askai.png	true	true	5	Setup	Rocket

Ask AI

Ask AI is a header button that opens a chat modal for asking questions about your site content. It uses RAG (Retrieval-Augmented Generation) to find relevant content and generate AI responses with source citations.

Press Cmd+J or Cmd+/ (Mac) or Ctrl+J or Ctrl+/ (Windows/Linux) to open the Ask AI modal.

---

How Ask AI works

+------------------+    +-------------------+    +------------------+
|  User question   |--->|  OpenAI Embedding |--->|  Vector Search   |
|  "How do I..."   |    |  text-embedding-  |    |  Find top 5      |
|                  |    |  ada-002          |    |  relevant pages  |
+------------------+    +-------------------+    +--------+---------+
                                                         |
                                                         v
+------------------+    +-------------------+    +------------------+
|  Streaming       |<---|  AI Model         |<---|  RAG Context     |
|  Response with   |    |  Claude/GPT-4o    |    |  Build prompt    |
|  Source Links    |    |  generates answer |    |  with content    |
+------------------+    +-------------------+    +------------------+

1. Your question is stored in the database with a session ID
2. Query is converted to a vector embedding using OpenAI
3. Convex vector search finds the 5 most relevant posts and pages
4. Content is combined into a RAG prompt with system instructions
5. AI model generates an answer based only on your site content
6. Response streams in real-time with source citations appended

Features

Feature	Description
Streaming	Responses appear word-by-word in real-time
Model Selection	Choose between Claude Sonnet 4 or GPT-4o
Source Citations	Every response includes links to source content
Markdown Rendering	Responses support full markdown formatting
Internal Links	Links to your pages use React Router (no page reload)
Copy Response	Hover over any response to copy it to clipboard
Keyboard Shortcuts	Cmd+J or Cmd+/ to open, Escape to close, Enter to send

Configuration

Ask AI requires semantic search to be enabled (for embeddings):

// src/config/siteConfig.ts
semanticSearch: {
  enabled: true,
},

askAI: {
  enabled: true,
  defaultModel: "claude-sonnet-4-20250514",
  models: [
    { id: "claude-sonnet-4-20250514", name: "Claude Sonnet 4", provider: "anthropic" },
    { id: "gpt-4o", name: "GPT-4o", provider: "openai" },
  ],
},

Environment variables

Set these in your Convex dashboard:

# Required for embeddings (vector search)
npx convex env set OPENAI_API_KEY sk-your-key-here

# Required for Claude models
npx convex env set ANTHROPIC_API_KEY sk-ant-your-key-here

After setting environment variables, run npm run sync to generate embeddings for your content.

When to use Ask AI vs Search

Use Case	Tool
Quick navigation to a known page	Keyword Search (Cmd+K)
Find exact code or commands	Keyword Search
"How do I do X?" questions	Ask AI (Cmd+J or Cmd+/)
Understanding a concept	Ask AI
Need highlighted matches on page	Keyword Search
Want AI-synthesized answers	Ask AI

Technical details

Frontend:

File	Purpose
src/components/AskAIModal.tsx	Chat modal with streaming messages
src/components/Layout.tsx	Header button and keyboard shortcuts
src/config/siteConfig.ts	AskAIConfig interface and settings

Backend (Convex):

File	Purpose
convex/askAI.ts	Session mutations and queries (regular runtime)
convex/askAI.node.ts	HTTP streaming action (Node.js runtime)
convex/schema.ts	askAISessions table definition
convex/http.ts	/ask-ai-stream endpoint registration
convex/convex.config.ts	persistentTextStreaming component

Database:

The askAISessions table stores:

• question: The user's question
• streamId: Persistent Text Streaming ID
• model: Selected AI model ID
• createdAt: Timestamp
• sources: Optional array of cited sources

Limitations

• Requires semantic search: Embeddings must be generated for content
• API costs: Each query costs embedding generation (~$0.0001) plus AI model usage
• Latency: ~1-3 seconds for initial response (embedding + search + AI)
• Content scope: Only searches published posts and pages
• No conversation history: Each session starts fresh (no multi-turn context)

Troubleshooting

"Failed to load response" error:

1. Check that ANTHROPIC_API_KEY or OPENAI_API_KEY is set in Convex
2. Verify the API key is valid and has credits
3. Check browser console for specific error messages

Empty or irrelevant responses:

1. Run npm run sync to ensure embeddings are generated
2. Check that semanticSearch.enabled: true in siteConfig
3. Verify content exists in your posts/pages

Modal doesn't open:

1. Check that askAI.enabled: true in siteConfig
2. Check that semanticSearch.enabled: true in siteConfig
3. Both conditions must be true for the button to appear

Resources

• Semantic Search Documentation - How embeddings work
• Convex Persistent Text Streaming - Streaming component
• Convex Vector Search - Vector search documentation