paperless-gpt LLM-OCR never runs - failing with connection errors

[Ashkenazzio]

I am trying to LLM-OCR my receipts since I'm getting poor native performance. I can't get it to work properly even though my setup seems to be legit and healthy:

Environment

• Host OS: Arch Linux

---

Docker Compose

services:
  broker:
    image: redis:7-alpine
    user: "1100:1100"
    restart: always
    volumes:
      - ./data/redisdata:/data

  db:
    image: postgres:16-alpine
    user: "1100:1100"
    restart: always
    environment:
      POSTGRES_DB: ${PAPERLESS_DBNAME}
      POSTGRES_USER: ${PAPERLESS_DBUSER}
      POSTGRES_PASSWORD: ${PAPERLESS_DBPASS}
    volumes:
      - ./data/postgres:/var/lib/postgresql/data

  paperless-ngx:
    build:
      context: .
      dockerfile: Dockerfile
    user: "1100:1100"
    restart: always
    depends_on:
      - broker
      - db
    ports:
      - "8000:8000"
    env_file: .env
    environment:
      # Core DB & timezone
      PAPERLESS_DBHOST: ${PAPERLESS_DBHOST}
      PAPERLESS_DBNAME: ${PAPERLESS_DBNAME}
      PAPERLESS_DBUSER: ${PAPERLESS_DBUSER}
      PAPERLESS_DBPASS: ${PAPERLESS_DBPASS}
      PAPERLESS_TIME_ZONE: Asia/Jerusalem

      # OCRmyPDF preprocessing
      PAPERLESS_OCR_CLEAN: clean-final
      PAPERLESS_OCR_DESKEW: "true"
      PAPERLESS_OCR_ROTATE_PAGES: "true"
      PAPERLESS_OCR_ROTATE_THRESHOLD: "30"

      # Languages & whitelist
      PAPERLESS_OCR_LANGUAGE: "heb+eng"
      PAPERLESS_OCR_USER_ARGS: >-
        {"tesseract_oem":1,"tesseract_pagesegmode":6,
         "tesseract_config":["tessedit_char_whitelist=אבגדהוזחטיכלמנסעפצקרשת־ 0123456789"]}

      # RTL post-processing
      PAPERLESS_CONSUMER_POST_PROCESSING: /usr/local/bin/bidi_reorder.sh

      # Redis connection (was trying localhost before)
      PAPERLESS_REDIS: redis://broker:6379/0

      # Admin & API
      PAPERLESS_ADMIN_USER: ${PAPERLESS_ADMIN_USER}
      PAPERLESS_ADMIN_PASSWORD: ${PAPERLESS_ADMIN_PASSWORD}
      PAPERLESS_API_TOKEN: ${PAPERLESS_API_TOKEN}
      PAPERLESS_SECRET_KEY: ${PAPERLESS_SECRET_KEY}
      PAPERLESS_ALLOWED_HOSTS: ${PAPERLESS_ALLOWED_HOSTS}
      PAPERLESS_PUBLIC_URL: ${PAPERLESS_PUBLIC_URL}

    volumes:
      - ./data/appdata:/usr/src/paperless/data
      - ./data/media:/usr/src/paperless/media
      - ./data/consume:/usr/src/paperless/consume
      - ./data/export:/usr/src/paperless/export
      - ./scripts:/opt/paperless/scripts:ro

  paperless-gpt:
    image: icereed/paperless-gpt:latest
    user: "1100:1100"
    restart: always
    depends_on:
      - paperless-ngx
    env_file: .env
    environment:
      # Connect to Paperless-ngx
      PAPERLESS_BASE_URL:   http://paperless-ngx:8000
      PAPERLESS_URL:      http://paperless-ngx:8000
      PAPERLESS_API_TOKEN: ${PAPERLESS_API_TOKEN}
      PAPERLESS_TOKEN: ${PAPERLESS_API_TOKEN}
      # if you ever enable HTTPS in Paperless, skip cert checks:
      PAPERLESS_INSECURE_SKIP_VERIFY: "true"
      PAPERLESS_PUBLIC_URL: ${PAPERLESS_PUBLIC_URL}
      OCR_AUTO:                     "true"

      # LLM/OCR settings
      LLM_PROVIDER: ollama
      LLM_MODEL: deepseek-r1:8b
      OCR_PROVIDER: llm
      VISION_LLM_PROVIDER: ollama
      VISION_LLM_MODEL: deepseek-r1:8b
      OLLAMA_HOST: http://host.docker.internal:11434

      # Tagging & logging
      MANUAL_TAG: paperless-gpt
      AUTO_TAG: paperless-gpt-auto
      LOG_LEVEL: info

    volumes:
      - ./data/prompts:/app/prompts
      - ./data/paperless-gpt-db:/app/db

.env

# Database (for both Paperless-ngx & your Postgres service)
PAPERLESS_DBHOST=db
PAPERLESS_DBNAME=paperless
PAPERLESS_DBUSER=paperless
PAPERLESS_DBPASS=paperless

# Admin credentials for Paperless-ngx UI
PAPERLESS_ADMIN_USER=admin
PAPERLESS_ADMIN_PASSWORD=paperless

# API & security
PAPERLESS_API_TOKEN=fc07efd64982c6784dc74bd2c728c2578d540eb0
PAPERLESS_SECRET_KEY=wD8fiRBin4tSSaN9Q0-CJh_GLd7UrlDEfmtOiklrNGBnd1t3qcBW0nJhpIk_sMbj6MY

# Host & URL settings
PAPERLESS_ALLOWED_HOSTS=localhost,paperless-ngx
PAPERLESS_PUBLIC_URL=http://localhost:8000

Observed Behavior

1. Startup logs show only the banner and “Worker 0 started.”

2. Immediately thereafter, repeated errors:
   HTTP request failed: Get "http://paperless-ngx:8000/api/documents/?tags__name__iexact=paperless-gpt-ocr-auto…": dial tcp …: connect: connection refused

3. Even once Paperless-ngx is healthy (responding to curl http://localhost:8000/api/documents/), no auto-OCR loop ever succeeds.

4. I tagged a document with paperless-gpt in the UI, restarted the GPT container, but still no LLM execution.

Questions

1. Why does the worker keep failing to connect, even when the API is healthy?

2. What’s the intended lifecycle for the manual-tag OCR path—should it pick up tags added after startup, and if so how?

Any pointers to get the LLM-OCR worker to actually process documents would be greatly appreciated. Thanks!