When a document is split into chunks for indexing, individual chunks can lose the thread — "it restarts automatically" is useless as evidence if the chunk never says what it is. Content enrichment fixes chunks at indexing time, before they're embedded, so each one stands on its own. It's configured per source, so you can enrich your critical reference content and leave blog posts on the standard pipeline.
Open an embedding under Wyregraf → All Embeddings and switch to the Retrieval Accuracy tab.
- The enrichment switches. All enrichment switches, including Atomic facts, are off by default. Whatever enrichment combination you enable shares one LLM call per chunk during indexing.
- Estimated indexing calls shows what a reindex will spend with the current switches, before you run it.
- Index status tells you whether the live index still matches these settings.
The switches #
- Disable summaries / Disable questions — Summaries & Questions stay on by default for compatibility; flip these to skip their per-chunk generation calls on sources that don't need them.
- Structure-aware chunking — split along headings instead of raw length, and prefix each chunk with its heading path ("Pricing → Refunds") for search. Free; falls back to the standard splitter when a document has no usable structure.
- Contextual blurbs — a short document-aware context note attached to each chunk, so "the third step" still matches searches about the procedure it belongs to.
- Coreference rewriting — rewrites ambiguous references ("it", "the plugin", "this setting") into explicit names in the text handed to the AI as evidence. Your original text is stored untouched and remains what citations show.
- Entity extraction — records the products, versions, and identifiers each chunk mentions, which powers the chat-side Entity Safety switch.
- Atomic facts — generates separately indexed single-fact entries and is off by default, so indexing stays fast and adds no provider cost until you turn it on for a source. Facts are kept only when their exact supporting phrase exists in the source. Atomic-fact extraction shares the enrichment LLM call, then each accepted fact adds an embeddings-provider input and one indexed row. The number therefore depends on how many supported facts the source yields; disable the switch when that extra index size and embeddings cost are not useful for the source.
Product codes, SKUs, and error codes are protected during enrichment — the AI never gets a chance to mistype them, and any chunk where protection can't be guaranteed is kept in its original form automatically.
Reindexing after changes #
Enrichment happens at indexing time, so changing any of these switches (or the source's prompts or models) only affects newly indexed content. When settings drift from the live index, the Index status readout flips to Reindex required — the old index keeps serving answers until you reindex from the Index tab, so nothing breaks in the meantime.
What to do next #
Pair enriched sources with the chat-side switches in Hybrid Search & Reranking — entity extraction in particular does its best work once Entity Safety is on.
