Northwind Vault

The recipe

Build your own with an agent

Use this as a blueprint, not a script — adapt the fields and copy to whatever you’re making. Each step is a real prompt you can hand to a coding agent like Claude Code with the Headless MCP server connected; the terminals show the tool calls the agent makes, and the code is what it produces.

1. 1

   Create the project & turn on semantic search

   Semantic search is a per-project switch. Have the agent create the project and enable it — from then on, every file you upload gets embedded and indexed by meaning.

   MCP · Headless

   # Create the project and switch on semantic search

   › claude "Create a project called Northwind Vault and enable semantic search on it"

   ✓ create_project { name: "Northwind Vault", slug: "demo-docs" }

   ✓ update_project { semanticSearchEnabled: true }

   → uploads will be embedded and searchable by meaning
2. 2

   Upload your documents

   There are no content types to model here. Point the agent at your files and it uploads them straight to the media library, where Headless embeds each one in the background.

   Claude Code

   # No content types — just files

   › claude "Upload our policies, runbooks, and reports — the PDFs and text files in ./docs"

   ✓ upload_media × 9 (filename, content, mimeType)

   → 9 documents uploaded, embedded, and indexed
3. 3

   Mint a read-only API key

   The portal searches from the browser, so create a public, read-only key under Project Settings → API Keys. Test it by asking a question in plain language, not keywords.

   TerminalCopy

   curl -s "https://headless.build/api/v1/demo-docs/media?semantic=stolen%20laptop&limit=3" \
     -H "X-API-Key: hls_live_your_read_only_key" | jq

4. 4

   Build the search UI

   Now the frontend. Semantic search is just the media endpoint with a `semantic` query parameter — Headless returns each hit with a relevance score and the exact passage that matched, so the UI can show why a result came back.

   Claude Code

   # A search box over the media library

   › claude "Write a single self-contained index.html: a search box that queries the media library semantically and renders each result with its matched passage and a relevance score"

   ✓ Write public/demos/demo-docs/index.html

   → 1 file · one endpoint · meaning-based search
   index.htmlCopy

   // Semantic search is just the media endpoint with a `semantic` query.
   async function search(q) {
     const params = new URLSearchParams({
       semantic: q,          // a natural-language question, not keywords
       threshold: "0.12",    // minimum relevance to return
       limit: "10",
     });
     const res = await api(`/media?${params}`);
     if (res.semanticSearchEnabled === false) return showEnableNotice();
   
     // Each hit says how well it matched, and the passage that matched.
     return res.data.map((asset) => ({
       title: asset.title,
       score: asset._similarity,
       passage: asset._matchedChunk?.text,
     }));
   }

5. 5

   Ship it

   Point it at any pile of documents and deploy. There’s no schema to maintain — the meaning of the text is the index. Drop in a new file and it becomes searchable within moments, no code change required.