AI Search Phase 2 DOCX Extraction

Use this prompt to implement deterministic DOCX text extraction for AI indexing only.

Scope

• Work in src/lib/search/ai/extraction; do not use src/lib/content-extraction.ts.
• Preserve existing text-like and PDF extraction behavior.
• Support .docx and application/vnd.openxmlformats-officedocument.wordprocessingml.document, including parameterized MIME values.
• Keep legacy .doc, ODT, RTF, Google-native formats, PPTX, XLSX, OCR, and embedded image extraction out of this phase.
• Use mammoth.extractRawText for plain text and paragraph breaks where Mammoth provides them.
• Skip images, charts, SmartArt, diagrams, tracked-change/comment metadata, and complex formatting preservation.

Implementation Notes

• Add mammoth if it is not already present. Mammoth 1.12 ships lib/index.d.ts; @types/mammoth is not published.
• Add a dedicated DOCX parser module with a typed parser error carrying sanitized messages, cause, and parser context.
• Route DOCX inputs through the DOCX extractor before generic UTF decoding.
• Keep unsupported document policies for legacy .doc, ODT, RTF, and Google-native formats.
• Use existing normalization and byte-limit enforcement.
• Route BackendIndexingRunner telemetry with extractionMethod: "docx".
• Use sanitized taxonomy messages for corrupt and encrypted DOCX failures.

Required Tests

• Valid DOCX with text.
• Empty DOCX returns empty extractor output.
• Corrupt DOCX fails with the DOCX parse taxonomy.
• Password-protected/encrypted DOCX fails or skips with the DOCX encrypted taxonomy.