Goal

Answer document questions from the exported source artefact, not from a hidden retrieval index. The runtime context supplies the absolute Dokument-Markdown path and, where available, Asset-Verzeichnis path for each attached document.

Workflow

1. Start with the supplied absolute Markdown path; never derive a path from a file ID or filename.
2. Read the heading structure with rg '^#{1,6} ' <document.md>.
3. Search literal terms with rg -n -i '<terms>' <document.md>.
4. Read the surrounding range with sed -n '<start>,<end>p' <document.md>.
5. For broad or full-coverage work, progress through document.md in bounded line windows and keep the covered ranges in your reasoning.
6. Follow assets/... references for tables or pictures whose visual content can change the answer. Resolve them relative to the export directory.

Artefact layout

Refined Docling artefacts are structured as:

<source-file-stem>__<file-key>/docling-<source-hash>/
├── document.md
└── assets/

An AnyDoc fast artefact is instead typically .../fast-<source-hash>/document.md. It can be read immediately but may have no assets and no page/BBox provenance. A refined artefact preserves Docling's reading order and can include visual assets. Never infer page or bounding-box provenance from Markdown markup. UI renderers obtain that information from the document-element graph data.

Rules

• Do not look for former document retrieval tools, Chroma collections, embedding IDs, or FileChunk nodes. They no longer exist; the exported Markdown file is the document index.
• Prefer the supplied absolute paths; do not guess storage paths or use file IDs as retrieval keys.
• Quote or cite Markdown line ranges and asset names when this helps a user verify the source.
• Do not claim visual facts from an asset filename or a placeholder alone.
• If a document artefact is missing, say so instead of inventing content from metadata.