Upload and process a PDF from a URL or a local file path (≤100MB). Processing typically takes 0-3 minutes depending on document size (estimate: 2 seconds per page). Returns the new document and its processing status; call get_document() to confirm it has finished before reading its content.

Primary document retrieval tool. After orienting with get_folder_structure() (when available) — and, for queries that may involve typed attributes, pairing it with get_metadata_schema() so you know which fields metadata mode could filter on — use this for all document-related questions. The bare call returns root-level sub-folders and documents; pass folder_id to drill into a sub-folder level by level. Use sort="relevance" + query for semantic ranking. Do NOT jump to search_documents() first — keyword search and metadata mode are escalation paths, only after browse_documents(sort="relevance") + the persistence retries have failed. browse_documents does not filter by attribute — for typed-attribute queries, escalate to search_documents(mode="metadata").

Escalation tool — never the first discovery step. Always start with browse_documents() (after get_folder_structure() when visible). Pick a mode: • mode="query" — keyword precision when browse missed a known doc by precise term, acronym, or file-name fragment. AND match with graded OR fallback. When results carry relevance (0–1, the chance the document helps answer the query) they are ranked by it, and in a small library a term found only in a document's text still finds it; otherwise they are sorted by upload date — judge relevance yourself. • mode="metadata" — structured filter / sort / aggregates over the declared schema. Use after browse missed a typed-attribute query, OR for counts / averages / extremes browse cannot answer. Call get_metadata_schema() FIRST to confirm field names and types — guessing risks silent zero-match.

Orientation step: show the folder hierarchy as a tree (like tree -d). Call this before browse_documents() to plan targeted retrieval — folder names reveal content domains (e.g. "Research", "Finance") and return folder IDs you pass to browse_documents(folder_id=…). Wait for this result before issuing browse_documents() or search_documents(). For folder_id="root" (default), returns the entire folder tree; for a specific folder_id, returns that subtree only.

List declared metadata fields (names, types, descriptions) so you can build metadata_filter / sort / aggregates / group_by for search_documents(mode="metadata"), or a metadata_filter that narrows mode="query". Call it only when the question filters by a typed attribute (year, status, category, …) or asks for counts / averages / extremes / distributions — not as a routine first step; ordinary discovery goes through browse_documents(). Call it WITHOUT folder_id first: one call returns the account-level fields plus every folder that declares its own, with folder ids — no need for get_folder_structure() beforehand. Pass a folder_id only to read one folder's effective schema (it falls back to the account-level one). Reuse the result for the rest of the conversation.

Check a document's processing status and metadata. Call this before get_document_structure() or get_page_content() to confirm the document is ready (status === "completed").

Extract a document's hierarchical outline (headers, sections, page references). REQUIRED for documents over 20 pages — call this first to locate relevant sections, then pass their page numbers to get_page_content(). Use the part parameter to iterate large outlines until pagination.has_more is false.

Extract page content from a processed document. Use tight, targeted page ranges — never the whole document at once. For documents over 20 pages, call get_document_structure() first to pick relevant sections. Embedded image paths in the response feed into get_document_image().

Retrieve an image from a document — pass an image_path from a get_page_content() response: either an embedded image (from a content block) or a full rendered page (from the page_images array, useful when a page is scanned or its meaning depends on visual layout). Returns the image as an MCP image content block (base64 data + MIME type). Do NOT repeat the raw base64 value in your text response.

Permanently delete documents and all associated data. Only invoke when the user explicitly names the documents AND confirms deletion. Returns results — one entry per requested document: { doc_name, status: "deleted" | "not_found" | "failed", error? }. Inspect each entry for per-document failures. This action is irreversible.