docket.example.toml

# docket settings. Pass --config PATH or set DOCKET_CONFIG; the docket CLI and
# docket-api also read ./docket.toml (importing the library does not). Environment variables override this
# file; command-line options and ProcessOptions override both.
# Every line below shows the built-in default; uncomment what you change.
# `docket config show` prints the effective values and where each came from.

[llm]
# 'ollama' (local or cloud Ollama) or 'openai' (any OpenAI-compatible Chat Completions
# API: Mistral, OpenAI, Azure OpenAI, vLLM, LM Studio). Env: DOCKET_LLM_PROVIDER.
# provider = "ollama"
# Ollama server URL. Env: OLLAMA_HOST.
# ollama_host = "http://localhost:11434"
# Base URL of the OpenAI-compatible API. Env: DOCKET_LLM_BASE_URL.
# base_url = "https://api.openai.com/v1"
# API key for the OpenAI-compatible API. Env: DOCKET_LLM_API_KEY.
# api_key = ""  # unset
# Model for classification and extraction (required). Env: DOCKET_TEXT_MODEL.
# text_model = ""
# Model that reads page images (the vlm OCR backend); empty makes vlm unavailable.
# Env: DOCKET_VISION_MODEL.
# vision_model = ""
# openai provider: 'json_schema' sends the extraction schema for constrained decoding
# (falling back to 'json_object' where the server refuses it); 'json_object' never
# sends it. Env: DOCKET_LLM_STRUCTURED_OUTPUT.
# structured_output = "json_schema"
# Extra attempts when the model's output fails the schema. Env:
# DOCKET_MAX_EXTRACT_RETRIES.
# max_extract_retries = 2
# Longest text sent in one extraction call; longer documents are chunked. Env:
# DOCKET_EXTRACT_CHUNK_CHARS.
# extract_chunk_chars = 12000
# Timeout for one vision-model page, seconds. Env: DOCKET_VISION_TIMEOUT_S.
# vision_timeout_s = 300.0
# Timeout for one request to a hosted model (Ollama cloud, a remote
# OpenAI-compatible API); local models keep their longer timeouts. Env:
# DOCKET_LLM_HOSTED_TIMEOUT_S.
# hosted_timeout_s = 60.0
# Retries after 429/5xx or a connection error, and after a timeout for hosted
# models. Env: DOCKET_LLM_RETRIES.
# retries = 2
# Let reasoning models think before answering. Env: DOCKET_ENABLE_THINKING.
# enable_thinking = false
# Process-wide limit on simultaneous LLM requests. Env: DOCKET_LLM_CONCURRENCY.
# concurrency = 4

[ocr]
# Primary OCR backend: tesseract, paddle, auto (first installed) or a plugin name. Env:
# DOCKET_OCR_BACKEND.
# backend = "auto"
# Backends tried in order when a page's reading is rejected (env: comma-separated). Env:
# DOCKET_OCR_FALLBACKS.
# fallbacks = ["vlm"]
# ISO 639-1 language codes; each extra language slows Tesseract. Env:
# DOCKET_OCR_LANGUAGES.
# languages = ["en"]
# A page reading is accepted at this confidence (0..1). Env: DOCKET_OCR_MIN_CONFIDENCE.
# min_confidence = 0.6
# Detect and undo page rotation on scans (Tesseract OSD). Env:
# DOCKET_OCR_DETECT_ROTATION.
# detect_rotation = true
# Correct fine scan skew before raster OCR. Env: DOCKET_OCR_DESKEW.
# deskew = true
# Find the document in a photo (receipt or invoice on a table) and flatten it
# before OCR; needs the [photo] extra, otherwise photos are read as they are.
# Env: DOCKET_OCR_CROP_PHOTOS.
# crop_photos = true
# Tesseract re-reads a page enlarged when its typical word is shorter than
# this many pixels (low-resolution scans); 0 disables. Env: DOCKET_OCR_MIN_TEXT_HEIGHT.
# min_text_height = 20.0
# Resolution PDF pages are rendered at for OCR. Env: DOCKET_OCR_DPI.
# dpi = 200
# Tesseract page-segmentation mode (3 automatic; 6 can suit dense tables). Env:
# DOCKET_TESSERACT_PSM.
# tesseract_psm = 3
# Process-wide limit on simultaneously running OCR engines (default: half the CPUs).
# Env: DOCKET_OCR_CONCURRENCY.
# concurrency = 4  # default: half the CPUs
# Ask a text model whether OCR text is garbled. Env: DOCKET_OCR_QUALITY_CHECK.
# quality_check = true
# Confidence the quality check needs to call text garbled. Env:
# DOCKET_OCR_QUALITY_MIN_CONFIDENCE.
# quality_min_confidence = 0.7

[paddle]
# PaddleOCR device: cpu, gpu or gpu:N. Env: DOCKET_PADDLE_DEVICE.
# device = "cpu"
# PaddleOCR model size. Env: DOCKET_PADDLE_MODEL.
# model = "mobile"
# Run PaddleOCR's table-structure pipeline. Env: DOCKET_PADDLE_TABLES.
# tables = false

[docling]
# TableFormer speed/quality mode. Env: DOCKET_DOCLING_TABLE_MODE.
# table_mode = "accurate"
# Match TableFormer cells back to recognized text. Env: DOCKET_DOCLING_CELL_MATCHING.
# cell_matching = true

[layout]
# Keep page layouts (words, lines, tables with coordinates) in results. Env:
# DOCKET_INCLUDE_LAYOUT.
# include = true
# Mark tables and columns ([TABLE n], [COLUMN n]) in the page text the LLM reads. Env:
# DOCKET_LAYOUT_MARKERS.
# markers = true

[classify]
# Confidence TF-IDF needs to answer instead of the LLM. Env:
# DOCKET_TFIDF_CONFIDENCE_FLOOR.
# tfidf_confidence_floor = 0.65
# Below this classification confidence a document goes to review. Env:
# DOCKET_MIN_CONFIDENCE.
# min_confidence = 0.55

[review]
# Persist flagged documents in the review queue. Disabled by default for the
# Python library; enable it explicitly for the review UI. Env:
# DOCKET_REVIEW_QUEUE_ENABLED.
# enabled = false
# A key field (total, date, number, parties) read from OCR words below this
# confidence sends the document to review. Env: DOCKET_MIN_SOURCE_CONFIDENCE.
# min_source_confidence = 0.8
# SQLAlchemy URL for SQLite or PostgreSQL review storage. Env: DOCKET_REVIEW_DATABASE_URL.
# database_url = "sqlite:///data/review.db"
# Review task lease duration before another reviewer can claim it. Env: DOCKET_REVIEW_LOCK_SECONDS.
# lock_seconds = 300
# Where copies of flagged documents are kept. Env: DOCKET_REVIEW_DOCUMENTS.
# documents = "data/review_documents"

[input]
# Largest accepted file. Env: DOCKET_MAX_FILE_BYTES.
# max_file_bytes = 20971520
# Largest decoded image or rendered PDF page in pixels. Env:
# DOCKET_MAX_IMAGE_PIXELS.
# max_image_pixels = 50000000
# Most pages read from one PDF. Env: DOCKET_MAX_PDF_PAGES.
# max_pdf_pages = 100

[batch]
# Documents in flight per batch; beyond the LLM limit they only help while others are in
# OCR. Env: DOCKET_BATCH_WORKERS.
# workers = 4

[api]
# Most files in one HTTP job. Env: DOCKET_MAX_BATCH_FILES.
# max_batch_files = 100
# Largest total upload of one HTTP job. Env: DOCKET_MAX_BATCH_BYTES.
# max_batch_bytes = 209715200
# HTTP jobs processed at once; the rest wait. Env: DOCKET_MAX_CONCURRENT_JOBS.
# max_concurrent_jobs = 2
# Bearer token the HTTP API requires when set. Env: DOCKET_API_KEY.
# key = ""  # unset
# Browser origins allowed to call the HTTP API. Env: DOCKET_API_CORS_ORIGINS.
# cors_origins = ["http://localhost:3000", "http://127.0.0.1:3000"]
# One directory per HTTP job: metadata, results, uploads until done. Env:
# DOCKET_JOBS_DIR.
# jobs_dir = "data/jobs"

[einvoice]
# Your own copy of the validation artifacts (default: the one in the package). Env:
# DOCKET_EINVOICE_RESOURCES.
# resources = ""  # unset
# Where `docket einvoice fetch` puts the artifacts Docket does not ship
# (Peppol, CII D16B, Factur-X; default: ~/.cache/docket/einvoice). Env:
# DOCKET_EINVOICE_DOWNLOADS.
# downloads = ""  # unset

[eval]
# Blended $/1M tokens used for the illustrative cost in eval reports. Env:
# DOCKET_CLOUD_COST_PER_1M_TOKENS.
# cloud_cost_per_1m_tokens = 0.2

[tracing]
# Langfuse tracing (optional). Env: LANGFUSE_PUBLIC_KEY.
# langfuse_public_key = ""  # unset
# Langfuse tracing (optional). Env: LANGFUSE_SECRET_KEY.
# langfuse_secret_key = ""  # unset
# Langfuse server. Env: LANGFUSE_HOST.
# langfuse_host = "https://cloud.langfuse.com"
# Send prompts and answers, i.e. the document text, to Langfuse. Off, traces
# carry model, latency and sizes only. Env: DOCKET_LANGFUSE_CONTENT.
# langfuse_content = false