Document pipeline spec template

Reference template for the document pipeline. Every encode strategy and every generate option is documented inline. Copy this into ~/.config/mm/pipelines/document/{mode}.yaml (replacing {mode} with fast or accurate) and edit only the fields you want to override — omitted keys fall back to the built-in defaults.

Source: python/mm/pipelines/document/spec.yaml.template

# Document pipeline spec — all available strategies and their options.
#
# Copy this file to ~/.config/mm/pipelines/document/{mode}.yaml
# and customise the values you need.  Omitted strategy_opts fall back
# to their defaults.  The generate block is optional — leave it out
# for encode-only (text extraction) pipelines.

kind: document
mode: fast        # fast | accurate

# ── encode ──────────────────────────────────────────────────────────

encode:

  # ── page-text (default) ─────────────────────────────────────────
  # Extracts text per page from PDF/DOCX/PPTX.  No rasterization —
  # fast and lightweight.
  strategy: page-text
  strategy_opts:
    pages_per_message: 128    # pages batched into each Message
    # max_pages: null         # total page cap (null = unlimited)

  # ── rasterize ───────────────────────────────────────────────────
  # Renders PDF pages as JPEG images.  Best when layout matters
  # more than raw text (charts, forms, diagrams).
  # strategy: rasterize
  # strategy_opts:
  #   max_width: 1024         # render width in pixels
  #   pages_per_message: 4    # pages per Message
  #   max_pages: null         # total page cap (null = unlimited)

  # ── rasterize-text ──────────────────────────────────────────────
  # Rasterized page images interleaved with extracted page text.
  # Gives the VLM both visual layout and raw text for OCR fallback.
  # strategy: rasterize-text
  # strategy_opts:
  #   max_width: 1024         # render width in pixels
  #   pages_per_message: 4    # pages per Message
  #   max_pages: null         # total page cap (null = unlimited)

  # ── gemini ─────────────────────────────────────────────────────
  # Passes the document directly as a Gemini inline_data Part.
  # No strategy_opts — the raw file is sent as-is.
  # strategy: gemini

  # Custom post-processing transform (optional).
  # pyfunc: null

# ── generate (optional — omit for encode-only) ──────────────────

generate:
  prompt: >-
    Structure this document ({filename}) into a clean markdown readable format.
    The conversion should be lossless and well-organized/formatted with proper
    headings, lists, etc.

    ```text
    {content}
    ```
  max_tokens: 16384
  # model: null              # pin a specific LLM model (overrides profile)
  # temperature: null        # sampling temperature (null = model default)
  # json_mode: false         # request JSON-formatted response
  # think: false             # enable extended thinking
  # reasoning_effort: none   # none | low | medium | high
  # extra_body: {}           # provider-specific pass-through kwargs