Manifest Reference

Field reference for manifest.jsonl chunk and deletion records. For behavior and configuration, see the Pipelines overview. To inspect a manifest interactively, see scripts/inspect-manifest.sh.

Record types

There are two record types:

Chunk record — a content section, ready to embed or index.

Deletion record — {"_deleted": true, "id": "..."} — a stale chunk ID to remove. Braised emits these for pages that were deleted or whose content changed since the last build.

Braised tracks source-file hashes in .braised/index-state.json. After the first build, only new or changed pages produce chunk records — the manifest is always incremental. On a clean first build every record is a chunk record. See Writing a Pipeline Script for how to handle deletion records and CI state.

Chunk record fields

{
  "id":            "/getting-started/install/#linux",
  "url":           "/getting-started/install/",
  "title":         "Installation Guide",
  "heading":       "Linux",
  "heading_level": 3,
  "breadcrumb":    ["Installation Guide", "Installation", "Linux"],
  "content":       "Installation Guide > Installation > Linux\n\nDownload the binary from the releases page...",
  "hash":          "a3f9c2b1"
}

Field	Type	Description
`id`	string	Chunk identifier. Format: `/url#anchor` for heading chunks, `/url` (no trailing slash) for page-root chunks. Note: page-root `id` drops the trailing slash that `url` carries. Renaming a heading changes its `id` — treat it as stable within a build, not across renames. Use as the vector DB document ID.
`url`	string	Output URL of the source page, with trailing slash.
`title`	string	Page title from frontmatter.
`heading`	string	Text of the nearest heading. Equals `title` for page-root chunks.
`heading_level`	int	Heading depth: `0` = page root (content before the first heading), `1`–`6` = H1–H6.
`breadcrumb`	[]string	Ancestor heading labels from page title down to the current heading. Always starts with `title`.
`content`	string	Pre-baked breadcrumb context followed by the plain-text chunk body. Pass this directly to your embedding model.
`hash`	string	FNV-32a hex digest of `content`, zero-padded to 8 hex characters. Used by braised for incremental change detection.

The `content` field

content is plain text — no Markdown, no HTML. Inline formatting is stripped. Braised prepends a breadcrumb context line automatically:

Installation Guide > Installation > Linux

Download the binary from the releases page and place it somewhere on your PATH.
Verify the installation by running: braised --version

Do not strip the breadcrumb prefix — it gives the model the hierarchical context it needs for accurate retrieval.

Page-root chunks

Content that appears before the first heading in a page becomes a chunk with heading_level: 0. The heading field equals the page title and the id has no fragment. These are valid embedding targets — introductory prose often carries useful semantic signal.

What is excluded

Draft pages (draft: true in frontmatter) are excluded from the build entirely and produce no chunks.
Pages that fail to build (missing title, broken frontmatter) are skipped and produce no chunks.
Empty sections — a heading with no content below it produces no chunk.
Pages with llm_exclude: true in frontmatter are excluded from all agent outputs including the manifest.