Vercel's AI agent skips the vector DB entirely

What it does

This is a starter kit for building AI agents that answer questions from your own knowledge base — documentation repos, YouTube transcripts, custom APIs — without ever chunking or embedding anything. The agent lives inside isolated sandboxes where it runs ordinary shell commands (grep, find, cat) against a cloned snapshot of your sources. Same agent code powers a web chat UI, a GitHub issue bot, and a Discord bot through pluggable adapters.

The interesting bit

The deliberate rejection of RAG orthodoxy. No vector database, no chunking pipeline, no embedding model to maintain. Search is deterministic, explainable, and instant — the trade-off being it only works when your knowledge fits in a filesystem and your questions match literal text. The “smart complexity router” is a nice touch: incoming questions are classified trivial-to-complex and routed to cheaper or pricier models automatically.

Key highlights

• File-based search via sandboxed shell commands — zero vector infrastructure
• Multi-platform deployment from one codebase: web chat, GitHub bot, Discord bot (Slack/Linear “soon”)
• Built-in admin panel with usage stats, error logs, source sync controls, and an AI-powered “admin agent” that can query its own telemetry
• Shared sandbox pool: warm sandboxes start in <100ms, cold spin-up in 1–3s
• AI-assisted customization via local “skills” in .agents/skills/ — delegate renaming, adding sources, or adding tools to an agent

Caveats

• Requires Vercel ecosystem lock-in: Vercel Sandboxes, Vercel Workflows, NuxtHub (SQLite), and the Vercel AI SDK
• “No embeddings” is a feature until your knowledge base needs semantic search across paraphrased concepts
• Several platform adapters (Slack, Linear) are listed as future work, not shipped

Verdict

Worth a look if you want a quick, opinionated agent that answers from docs and issues where literal text search suffices — and you already live in Vercel’s stack. Skip it if you need semantic retrieval, multi-tenant isolation, or freedom from platform-specific infrastructure.