AI Document Engine

Use the File AI MCP tools as the source of truth for document content. Do not manually unzip Office files, inspect PDF internals, or infer spreadsheet structure when these tools are available.

File AI format recognition follows the Flyfish File Viewer core registry. Treat profile.format as the authoritative format decision and profile.extraction as the current content-awareness strategy for that format.

Core Workflow

1. Call doc_analyze with the local file path before answering content questions.
2. Use profile.indexId for all later calls in the same task.
3. Use doc_context when the user asks about a topic, section, page, slide, sheet, or nearby content.
4. Use doc_search for exact terms, entities, clauses, headings, table labels, filenames inside archives, or attachment names.
5. Use doc_read for a specific anchorId, blockId, or chunkId returned by the engine.
6. Cite or mention the relevant anchorId values in answers when making document-grounded claims.

Tool Rules

• Prefer doc_context over reading all chunks when the user asks a focused question.
• Prefer doc_search before broad summarization if the user provides keywords or entity names.
• Use doc://{indexId}/profile, doc://{indexId}/outline, and doc://{indexId}/chunks resources only after doc_analyze has cached the document.
• Treat warnings in profile.warnings as part of the evidence, especially scanned PDFs, clipped sheets, unsupported formats, or best-effort fallback text.
• Use doc_list_formats when choosing whether a file extension is recognized by File Viewer; do not infer supported formats from extractor names alone.
• If profile.format.known is true but profile.extraction.structured is false, say the file is recognized by File Viewer but File AI only has metadata or best-effort text for that renderer-only format.
• For spreadsheets, keep sheet names and row context in the answer when available.
• For presentations, keep slide numbers in the answer when available.
• For PDFs, keep page numbers in the answer when available.
• For archives, report manifest entries and nested paths; do not extract or mutate archive entries unless another explicit tool supports it.

Response Discipline

Ground answers in returned anchors and excerpts. If the engine reports missing extractable text or clipped content, say so briefly and avoid overclaiming. This skill is read-only; do not promise edits, conversions, or document rewrites through this MCP server.