Build knowledge agents without embeddings
By Steven Van ·
Vercel's open-sourced Knowledge Agent Template swaps vector databases for grep and bash in a Sandbox, cutting one internal agent's cost by 75%.
Vercel's sales call summarization agent used to cost about $1.00 per call. After swapping its vector search pipeline for a plain filesystem and giving the agent bash access, the cost dropped to about $0.25 per call, and the output quality improved. Vercel says the problem with embedding-based retrieval is that failures are silent: the agent confidently returns the wrong chunk, with no way to trace how it got there.
Vercel has open-sourced that architecture as the Knowledge Agent Template, built on Vercel Sandbox, AI SDK and Chat SDK. Instead of a vector database, a chunking pipeline and an embedding model, the agent runs grep, find and cat against a snapshot of the source content inside an isolated Sandbox. Sources are added through an admin interface and stored in Postgres, then synced to a snapshot repository via Vercel Workflow; when the agent needs to search, a Sandbox loads that snapshot and its bash tools query it directly. Vercel says this makes wrong answers traceable: opening the trace shows which command the agent ran, such as grep -r "pricing" docs/, and which file it read.
- Sources can be GitHub repos, YouTube transcripts, documents such as markdown files, or custom APIs.
- The template can be shipped as a web chat app, a GitHub bot, a Discord bot, or all three at once.
- It deploys to Vercel in a single click.
