Creating Embeddings

An Embedding is a batch of content turned into Wyregraf's searchable AI index. You pick a content source — your posts, a PDF, a public URL — and Wyregraf splits it into chunks, sends each chunk to your embeddings provider, and stores the resulting vectors in your WordPress database. Chats then answer questions by searching those chunks.

Create an embedding #

Go to Wyregraf → All Embeddings.

1. Add New starts a new embedding.
2. Each row is one indexed batch; the same content can be re-indexed from here at any time after it changes.

Give the embedding a title, then work through the editor's tabs.

The Index tab #

1. Content Source picks where the text comes from:

   • Posts — content from the post types you enabled in the embeddings profile settings.
   • Internal — text you paste directly into the embedding.
   • Code, PDF, and URL Scrape (Pro) — index a zip of code files, an uploaded PDF, or any public web page by URL. Each has its own options — see Indexing PDFs, URLs & Code.

   Below the source, Wyregraf shows the token count and estimated cost of indexing, so you know what an index run will spend before you start it.

2. Create Embeddings queues the content for indexing; the queue processes in the background and shows its progress here. Once indexed, the button becomes Reindex — run it again whenever the content changes.

3. View Embeddings opens the chunks viewer.

Reviewing the chunks #

The chunks viewer shows exactly what was indexed — every chunk with its parent reference, searchable by text and filterable by type (standard, summary, questions).

Skim a few chunks after your first index run. If chunks cut mid-thought or bundle unrelated topics, adjust the chunk size in the embeddings profile and reindex.

The Metadata tab #

1. Document Title and the other metadata fields describe the source. They are what a visitor sees when a chat cites this content, so give them the reader-facing name rather than an internal label.
2. Source Origin is the URL a citation links back to.

The Profiles tab #

Each embedding uses your default Embeddings & Vector profile and LLM profile; the Profiles tab lets a specific embedding override that — for instance to index one document with a larger-chunk profile.

The Retrieval Accuracy tab #

The Retrieval Accuracy tab holds this source's enrichment and chunking switches, an Estimated indexing calls readout so you can see what a reindex will spend before running it, and an Index status line that flips to Reindex required whenever a setting change means the live index no longer matches — the old index keeps serving until you reindex.

What to do next #

With content indexed, create a chat that answers from it — or first add Summaries & Questions to sharpen what the search can find.