A chat widget answering questions from your own documentation is a well-documented pattern (retrieval + an LLM call). Handing off to a human cleanly, and not confidently making things up, is where the real engineering care goes.
Borderline
Effort: a weekendRunning cost: $5–$10/mo
Why CanICodeThis says Borderline
A website chat widget with real-time messages is the same buildable core as any chat app on this register
Retrieval-augmented generation (search your own docs, feed relevant chunks to an LLM) is a well-documented, mature pattern now
Logging conversations for later review is ordinary CRUD
The parts that'll cause problems
Keeping the bot from confidently inventing wrong answers (hallucination) needs real prompt and retrieval discipline, not just 'add an LLM call'
Clean handoff to a human agent when the bot can't help is a genuine, separate workflow feature
Multi-channel support (email, SMS, not just website chat) is real integration work per channel
How I'd build this
frontend
Next.js
database
Supabase (with pgvector for retrieval)
hosting
Vercel / Cloudflare Pages
MVP scope
A website chat widget with real-time messages
Answers generated by retrieving relevant docs and calling an LLM
Conversation logging for review
A clear 'talk to a human' escape hatch
Postpone to v2
Email and SMS channels, not just website chat
Analytics on what people ask and where the bot fails
Fine-tuned handoff routing to specific team members
What it actually costs to build
AI agent — with a subscription (Claude Pro/Max, Cursor, etc.)
$0 marginal
AI agent — pay-per-use API, no subscription
$43–$86 one-time
Hosting, once it's running
$5–$10/mo
Domain name, if you want your own
~$12/yr
Existing tools solving a similar problem typically run $39–$55/mo — see how they compare below.
The build prompt
Build me an AI support chat widget — genuinely achievable with RAG, not a custom model:
1. A chat bubble embedded on your site, real-time over WebSockets.
2. Index your documentation into a vector database (Supabase with pgvector, or Pinecone) — this is the retrieval half of RAG.
3. On each message, retrieve the most relevant doc chunks and call an LLM API to generate an answer grounded in them, not the model's general knowledge.
4. Always show a clear 'talk to a human' option, and log every conversation for review.
That's a real, working support bot. Out of scope: email and SMS channels beyond website chat, real per-channel integration work, and fine-grained handoff routing to specific team members, which needs real routing logic beyond a single escalation path.