# Copy to .env and fill in. .env is git-ignored.
# Every commented-out line shows the default. See README.md > Configuration.

# --- Required -----------------------------------------------------------------
MISTRAL_API_KEY=your-mistral-api-key

# --- Docker compose -----------------------------------------------------------
# Host folders mounted at /data (inputs) and /config (work folder).
DATA_PATH=./data
CONFIG_PATH=./config
# linuxserver.io conventions: user and group that own the data folder, the
# permission mask for new files (002: group-writable), and the time zone.
PUID=1000
PGID=1000
UMASK=022
TZ=Etc/UTC

# --- Inputs and folders -------------------------------------------------------
# stacks/ splits large scans into documents; scanner/ takes one document per
# file (e.g. a scanner saving to a network share). Each has inbox/, output/,
# archive/ and failed/ below its root.
# STACKS_ENABLED=true
# SCANNER_ENABLED=true
# DATA_DIR=/data
# Text for splitting and naming: mistral (Mistral OCR, paid, splits somewhat
# better) or tesseract (the text layer ocrmypdf adds anyway, free, no wait).
# See README > Choosing the text source.
# STACKS_TEXT_SOURCE=mistral
# Default for the scanner: tesseract (mistral if OCRMYPDF_ENABLED=false).
# SCANNER_TEXT_SOURCE=tesseract
# STACKS_DIR=/data/stacks
# SCANNER_DIR=/data/scanner
# WORK_DIR=/config
# Days after which a processed file's work folder (searchable copy, OCR text,
# plan.json) is deleted. rebuild works until then. 0 keeps everything.
# WORK_RETENTION_DAYS=30
# Temporary page images; default: $WORK_DIR/tmp, on disk rather than in RAM.
# TMPDIR=
# A file must stay unchanged this long (seconds) before it is picked up.
# STABLE_SECONDS=60
# POLL_INTERVAL=30

# --- Paperless-ngx input -----------------------------------------------------
# Setting PAPERLESS_URL enables a third inbox (paperless/inbox/). Its files get
# the Tesseract text layer and are uploaded to Paperless-ngx, which names and
# tags them itself.
# PAPERLESS_URL=http://paperless:8000
# PAPERLESS_TOKEN=your-paperless-api-token
# PAPERLESS_DIR=/data/paperless
# Comma-separated tag ids to add on upload.
# PAPERLESS_TAGS=
# PAPERLESS_MAX_WAIT_MINUTES=30
# mistral: replace the document's content in Paperless with Mistral OCR's text
# (tables kept as Markdown). tesseract: keep the text layer's plain text.
# PAPERLESS_TEXT_SOURCE=tesseract
# A second Paperless input (paperless-2/inbox/) that uploads with its own
# token, so its documents belong to another Paperless user. Without
# PAPERLESS_2_URL it uses the same instance as PAPERLESS_URL.
# PAPERLESS_2_TOKEN=api-token-of-the-second-user
# PAPERLESS_2_URL=
# PAPERLESS_2_DIR=/data/paperless-2
# PAPERLESS_2_TAGS=
# PAPERLESS_2_TEXT_SOURCE=tesseract
# Make every tag, correspondent or document type ownerless, so all Paperless
# users see and may change it. Checked every PAPERLESS_SHARE_TAGS_MINUTES.
# Needs a token that may change other users' objects (a superuser's). Tag
# names in PAPERLESS_SHARE_TAGS_READONLY keep their owner and are only
# visible to everyone else.
# PAPERLESS_SHARE_TAGS=false
# PAPERLESS_SHARE_CORRESPONDENTS=false
# PAPERLESS_SHARE_DOCUMENT_TYPES=false
# PAPERLESS_SHARE_TAGS_MINUTES=1
# PAPERLESS_SHARE_TAGS_READONLY=ai-processed

# --- Mistral ------------------------------------------------------------------
# MISTRAL_LLM_MODEL=mistral-large-latest
# MISTRAL_OCR_MODEL=mistral-ocr-latest
# Requests per second across all workers. Your per-model limit is shown in
# Mistral's console under API > Limits; 0 disables the throttle.
# MISTRAL_MAX_RPS=1
# MISTRAL_API_BASE=https://api.mistral.ai/v1
# MISTRAL_TIMEOUT=300

# --- OCR ----------------------------------------------------------------------
# batch: Mistral's batch API at half price; a job takes minutes, not seconds.
# direct: full price, immediate.
# OCR_MODE=batch
# BATCH_POLL_SECONDS=15
# A job still running after this long is cancelled; its chunks are retried directly.
# BATCH_MAX_WAIT_HOURS=24

# --- Spending limit -----------------------------------------------------------
# When Mistral refuses the account (spending limit, quota, rejected key),
# processing pauses and one file is retried as a probe this often.
# PAUSE_RETRY_MINUTES=30

# --- Naming -------------------------------------------------------------------
# Language for titles and summaries. Default: each document's own language.
# TITLE_LANGUAGE=English
# {title} is required, {date} is YYYY-MM-DD.
# FILENAME_PATTERN={title} {date}
# NO_DATE_LABEL=undated
# Text per document sent for naming; longer documents are shortened in the middle.
# METADATA_MAX_CHARS=24000

# --- Splitting and blank pages ------------------------------------------------
# Pages per model request, and how far each window moves on.
# BOUNDARY_WINDOW=12
# BOUNDARY_STEP=6
# Documents with a split decision below this confidence are flagged in review.md.
# REVIEW_CONFIDENCE=0.75
# DROP_BLANK_PAGES=true
# Pages with less visible ink than this (percent of the page) count as blank.
# BLANK_MAX_INK_PERCENT=0.2
# Pages without images and with at most this many characters count as blank.
# BLANK_MAX_CHARS=15

# --- Text layer (ocrmypdf / Tesseract) ----------------------------------------
# OCRMYPDF_ENABLED=true
# Tesseract languages, joined with +. deu, eng and osd are built in; others
# (e.g. fra, ita, spa) are downloaded once at startup into /config/tessdata.
# OCRMYPDF_LANGUAGES=deu+eng
# Download source for those, e.g. a local mirror of tessdata_best 4.1.0.
# TESSDATA_URL=https://github.com/tesseract-ocr/tessdata_best/raw/4.1.0
# Parallel OCR pages. auto: as many as the container's memory allows
# (1 GB base + 0.75 GB per page, at most one per CPU core). Less memory makes
# processing slower, not fail. Shared by all inputs; scanner and Paperless
# files go ahead of stacks.
# OCRMYPDF_JOBS=auto
# Images above this resolution are downsampled before OCR (0 = never).
# OCRMYPDF_MAX_IMAGE_DPI=600
# OCRMYPDF_EXTRA_ARGS=
# Limits, so one odd page can't exhaust the server. Peak memory is about
# OCRMYPDF_JOBS x 16 bytes x OCRMYPDF_MAX_OCR_MPIXELS (megapixels).
# OCRMYPDF_MAX_OCR_MPIXELS=50
# OCRMYPDF_PAGE_TIMEOUT=300
# OCRMYPDF_FILE_TIMEOUT_MINUTES=120
# OCRMYPDF_SKIP_BIG_MPIXELS=200

# --- Queue webhook ------------------------------------------------------------
# POSTs queue counts (JSON, no file names) on every change and as a heartbeat,
# e.g. to a Home Assistant webhook trigger. Empty disables it.
# QUEUE_WEBHOOK_URL=http://homeassistant.local:8123/api/webhook/your-random-id
# QUEUE_WEBHOOK_CHECK_SECONDS=10
# QUEUE_WEBHOOK_HEARTBEAT_SECONDS=300

# --- Throughput and logging ---------------------------------------------------
# OCR_CHUNK_PAGES=50
# OCR_CONCURRENCY=3
# LLM_CONCURRENCY=4
# LOG_LEVEL=INFO