Parse providers

Parsing is the first phase of training: Pegasus extracts text and structure from your raw documents and converts them to Markdown. The bot's vector index is built from that Markdown. Different parsers handle the same document differently — choose the one that fits your content.

Pegasus offers two parsers#

The choice lives in the Source settings tab as a "Parse Provider" selector.

Parse provider selector

When to pick Marker#

• The document is a scanned PDF (image-based) and you need accurate text.
• The document has dense tables, formulas, or figures that carry meaning.
• The content is sensitive and you prefer keeping parsing on the local pipeline.

Expect training to take noticeably longer.

When to pick LlamaParse#

• The document is text-native (exported from Word, generated from a template, or already Markdown).
• You want training to finish as quickly as possible.
• The visual layout is simple — paragraphs and headings.

Changing the provider mid-life#

You can switch the provider before each training. The choice applies to the next training only; previous parsed outputs are not retroactively re-parsed.

What happens if a parse fails#

If the parser fails on a file, training reports an error and the bot returns to its previous status. The bot's "Last error" field shows a brief reason. Try the other parser, or simplify/split the offending document.