doc-parse

Multi-format Document Parsing Skill (Mijuan Room)

Usage

• Invoked by the Mijuan Room Worker when it receives a file that has passed the security scan.
• Three-level fallback is attempted in sequence, outputting the final structured result and confidence.

Input

• file_path: local file path that has passed security-scan
• formats: (optional) whitelist of expected supported formats, defaults to all formats

Output

• markdown: structured Markdown body
• metadata: metadata such as title, page count, table count, author, etc.
• confidence: 0–1 parsing confidence
• fallback_used: name of the parser that ultimately took effect

Dependencies

• MinerU (preferred, strong layout restoration)
• PyMuPDF (second choice, fast PDF parsing)
• pdfplumber (fallback, table/text extraction)
• python-magic(type detection)

Failure Handling

• Primary parser fails → automatically degrade to the next tier, until all have been tried.
• All three tiers fail → mark 人工介入, do not output incomplete results, return BLOCKED to the Junjichu.
• Oversized single-page file → parse in chunks and then concatenate, to avoid memory overflow; a chunk-level failure only marks that chunk as low confidence.

Reuse Value

• General-purpose parsing foundation: RAG indexing, enterprise knowledge bases, and contract structuring can all be reused directly.
• Transparent confidence: downstream (Tongzhengsi four-color cards) can decide based on it whether manual review is needed, reducing hallucination risk.

Semifinal code package execution (runnable package)

• Real entry point:scripts/run_doc_parse.py
• Execution is equivalent to core.runtime.AgentSession.run_stage("doc-parse"), calling archive.mijuanfang.MiJuanFangAgent(three-level parsing fallback, pure Python and offline-capable).
• Run:python skills/doc-parse/scripts/run_doc_parse.py
• Artifacts:examples/snse_survey/skill_outputs/doc_parse.json(parsing result + confidence + fallback information).