Read a filing's content by `document_id` from `list_filings`. Numbers and prose live inside the document; `list_filings` metadata only locates filings.
RESPONSE SHAPES:
• `kind='embedded'` (under `max_bytes`, ~20 MB default) — full `bytes_base64`, `source_url_official` (evergreen registry URL), `source_url_direct` (short-TTL signed proxy URL). PDFs render as a native document block.
• `kind='resource_link'` (oversized) — NO `bytes_base64`. Returns `reason`, `next_steps`, both source URLs, and `index_preview` `{page_count, text_layer, outline_present}`. Call `get_document_navigation` to locate pages, then re-call this tool with `pages='N-M'` and `format='pdf'|'text'|'png'`.
When NOT to use: to enumerate a company's filings, use `list_filings`. To check size or available formats before deciding, use `get_document_metadata`. Never synthesize `document_id` — composite IDs will 404.
CRITICAL: on failure (rate limit / 5xx / timeout) do NOT fabricate names, numbers, or dates — tell the user what failed and offer retry or `source_url_official`. Outline titles, previews, and navigation snippets are for LOCATING pages, never for quoting.
`max_bytes` is a hard inline cutoff: raising it forces full proxy/R2 transfer (slower, costlier); the default returns `resource_link` for big PDFs so you can page-fetch. `fresh=true` bypasses the R2 cache and refetches from upstream — filings are immutable so it's rarely needed. `source_url_official` auto-resolves from the most recent `list_filings` call; `company_id` / `transaction_id` / `filing_type` / `filing_description` are overrides only when `document_id` did NOT come through `list_filings`.