Structured Extraction

Structured Extraction

Use this skill when a task depends on facts contained in uploaded documents or other files and those facts must be extracted before analysis.

Workflow

1. Read the ## Current attachments section in the system prompt. It gives each original filename plus its preprocessed Markdown and original-file paths.
2. For each source, call document_conversion_api_extractdocument directly with run_python only when the preprocessed Markdown is missing or a fresh extraction is explicitly needed. Pass the original path as file_path.
3. In that Python call, serialize the complete result to /home/user/work/<source-name>.json instead of printing it. For example:

import json

result = document_conversion_api_extractdocument(
    file_path="/home/user/attachments/<attachment-id>/original/<source-file>"
)
with open("/home/user/work/<source-file>.json", "w") as extracted:
    extracted.write(json.dumps(result, ensure_ascii=False, indent=2))
print("Saved extraction to /home/user/work/<source-file>.json")

4. Prefer reading or searching the preprocessed /home/user/attachments/<attachment-id>/content.md with read_file or run_bash commands such as grep; it remains available across later turns. Work files also persist without appearing as generated files in the conversation.
5. Preserve the source filename and every available location marker when reporting evidence. Keep extraction separate from interpretation.
6. If conversion fails or a format is unsupported, report the exact failure and do not substitute guessed content.

Xberg supports many document families, including PDF, Word and other office documents, spreadsheets, presentations, ebooks, email, archives, HTML/XML, structured text, images with OCR, and audio or video transcription. Let the conversion response determine whether a particular file is supported.

Quality checks

• Every extracted claim has a source filename and location when available.
• Tables retain their headers and row or column meaning.
• Empty, unreadable, or missing content is marked as unavailable.
• Extraction and interpretation remain separate.