I recently upgraded the search on my website to use AgentDB instead of a JSON file with the embeddings and vibe-coded dot products.

Try it out:

The search feature had been a broken and unnoticed for a while. I have very little tracking or and zero error monitoring in place, so when the api key to make the embeddings had stopped working I was unaware anything was amiss.

When I eventually noticed and got around to fixing it, I decided that the dead-simple JSON file setup was perhaps a bit too gritty and needed to be brought closer to something I might actually build and tell people about.

Now that it’s easier than ever to add low-effort features, it felt like a good time to give the search backend and upgrade.

Enter AgentDB.

AgentDB was a good fit for my needs: easy to host, difficult to break, generous free usage limits, and support for vector search.

Since we live in an age of wonders, the migration path was easy. I connected the AgentDB MCP server to Augment and asked politely for it to draft a schema and then create the table. Then I had Augment migrate the old embeddings, update the generation script, and change the search implementation to use the new table.

Setup

• create a uuid to namespace the table

• generate a schema and create tables

• import existing embeddings into d

  b

• query the db using vec_distance_cosine() to filter results

It was straightforward to replace the crude cosign similarity against the JSON file with a database query that uses a native cosign distance. At the pace I update my website, I’ll be able to operate comfortable for decades before I need to worry about storage or compute for similarity searches.