Plain chunk search has a blind spot: a visitor's question ("how do I speed up my site?") often shares no words — and not even much phrasing — with the passage that answers it. Wyregraf closes that gap with two extra embedding types built on top of your content:
- Summaries — an AI-written summary of each document, indexed alongside the original chunks. Broad questions match the summary even when no single chunk covers the whole topic.
- Questions — AI-generated questions that each document answers, indexed as searchable entries. A visitor's question matches a stored question far more reliably than it matches prose.
Either way, when a summary or question chunk matches, the citation shown to the visitor resolves to the original source document — readers always land on your real content, never on the generated helper text.
Both are optional. A chat works fine on standard chunks alone; add these when you want retrieval to catch broader or more conversational questions.
The Summary tab #
Open an embedding from Wyregraf → All Embeddings and switch to the Summary tab.
The Summary Prompt is the instruction the AI follows when writing the summary — the default works well, or tailor it (for example, "summarize for a beginner audience"). Create the summary embeddings from this tab's queue controls; like standard chunks, they process in the background and report how many chunks were indexed. Long summaries are automatically split to respect your profile's chunk size, staying linked to their parent document.
The Questions tab #
The Questions tab works the same way: a Questions Prompt controls how questions are generated from the content, and the tab's own queue creates and counts the question embeddings.
Cost and upkeep #
Summaries and questions are generated by your LLM profile's model and then embedded — one generation call per retrieval chunk each — so an index run with them enabled makes more API calls than a standard run. The Estimated indexing calls readout on the source's Retrieval Accuracy tab shows the exact count before you reindex. When the underlying content changes, reindex the summary and question embeddings along with the standard ones so all three stay in step.
Both are on by default. To skip them on a source that doesn't need them, flip Disable summaries or Disable questions on the same Retrieval Accuracy tab — see Content Enrichment & Chunking for everything else that tab controls.
