Some feedback trying to get this installed

[hughesadam87]

Hi - in case useful, here is my experience trying to get nv-ingest python quickstart guide working.

1. Got API keys - this was easy.
2. Following Quickstart Librfary Mode, I overlooked that linux env is required and hit dead end trying to install on windows.
3. Conda is falling out of favor in many orgs due to confusing licensing/product offering - I was able to get a conda-forge miniconda working, but it set me back a while. Would much prefer being able to build this with pip, poetry or just a pre-configured docker image has all the client library dependencies.

Here are the Docker/docker-compose files I settled on. Running conda in docker is also awkward IMO - also I had to figure out these system depenencies like libgl1-mesa-glx from chatgpt because I was getting library errors. Assuming I'm using the wrong base image?

# Use the official Miniconda image as a base
FROM continuumio/miniconda3:latest

# Install system dependencies required by OpenCV (libGL)
USER root
RUN apt-get update && apt-get install -y \
    libgl1-mesa-glx \
    && apt-get clean && rm -rf /var/lib/apt/lists/*

# Disable SSL verification for conda (not recommended for production)
RUN conda config --set ssl_verify false

# Create a new conda environment named "nvingest" with Python 3.10
RUN conda create -n nvingest python=3.10 -y && conda clean -afy

# Change the shell so subsequent commands run in the new environment
SHELL ["conda", "run", "-n", "nvingest", "/bin/bash", "-c"]

# Install NV-Ingest packages from the specified channels, one at a time
RUN conda install -y -c rapidsai -c conda-forge -c nvidia nv_ingest=25.3.0 && conda clean -afy
RUN conda install -y -c rapidsai -c conda-forge -c nvidia nv_ingest_client=25.3.0 && conda clean -afy
RUN conda install -y -c rapidsai -c conda-forge -c nvidia nv_ingest_api=25.3.0 && conda clean -afy

# (Optional) Install additional pip dependencies if needed.
RUN pip install opencv-python \
      llama-index-embeddings-nvidia \
      pymilvus \
      "pymilvus[bulk_writer, model]" \
      milvus-lite \
      nvidia-riva-client \
      unstructured-client

# Set the working directory inside the container
WORKDIR /app

# Copy your application code to the container
COPY . /app

# Set a default command to run your main.py file with the nv-ingest environment
CMD ["conda", "run", "-n", "nvingest", "python", "main.py"]

services:
  nvingest:
    build:
      context: .
      dockerfile: Dockerfile
    image: nvingest:latest
    container_name: nvingest
    command: ["conda", "run", "-n", "nvingest", "python", "main.py"]

For reference, the main.py file I was using from quickstart guide does not have the milvus portion. I'm only interested in extraction.

import logging, os, time, sys

from nv_ingest.util.pipeline.pipeline_runners import start_pipeline_subprocess
from nv_ingest_client.client import Ingestor, NvIngestClient
from nv_ingest_client.message_clients.simple.simple_client import SimpleClient
from nv_ingest.util.pipeline.pipeline_runners import PipelineCreationSchema
from nv_ingest_client.util.process_json_files import ingest_json_results_to_blob

# Start the pipeline subprocess for library mode                       
config = PipelineCreationSchema()                                                  

pipeline_process = start_pipeline_subprocess(config)
# you can configure the subprocesses to log stderr to stdout for debugging purposes
#pipeline_process = start_pipeline_subprocess(config, stderr=sys.stderr, stdout=sys.stdout)

client = NvIngestClient(
    message_client_allocator=SimpleClient,
    message_client_port=7671,
    message_client_hostname="localhost"
)

# Note: gpu_cagra accelerated indexing is not yet available in milvus-lite
# Provide a filename for milvus_uri to use milvus-lite
milvus_uri = "milvus.db"
collection_name = "test"
sparse=False

# do content extraction from files                                
ingestor = (
    Ingestor(client=client)
    .files("data/multimodal_test.pdf")
    .extract(              
        extract_text=True,
        extract_tables=True,
        extract_charts=True,
        extract_images=True,
        paddle_output_format="markdown",
        extract_infographics=True,
        # Slower, but maximally accurate, especially for PDFs with pages that are scanned images
        #extract_method="nemoretriever_parse",
        text_depth="page"
    )
)

print("Starting ingestion..")
t0 = time.time()
results = ingestor.ingest(show_progress=True)
t1 = time.time()
print(f"Time taken: {t1-t0} seconds")

# results blob is directly inspectable
print(ingest_json_results_to_blob(results[0]))

And it fails with import errors. Am I using the wrong base image for example?

Traceback (most recent call last):
2025-03-31T18:27:03.249284880Z   File "/app/main.py", line 3, in <module>
2025-03-31T18:27:03.249289085Z     from nv_ingest.util.pipeline.pipeline_runners import start_pipeline_subprocess
2025-03-31T18:27:03.249291551Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/util/pipeline/__init__.py", line 5, in <module>
2025-03-31T18:27:03.249293985Z     from .pipeline_builders import setup_ingestion_pipeline
2025-03-31T18:27:03.249296000Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/util/pipeline/pipeline_builders.py", line 10, in <module>
2025-03-31T18:27:03.249298468Z     from nv_ingest.util.pipeline.stage_builders import *
2025-03-31T18:27:03.249300406Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/util/pipeline/stage_builders.py", line 21, in <module>
2025-03-31T18:27:03.249302494Z     from nv_ingest.stages.docx_extractor_stage import generate_docx_extractor_stage
2025-03-31T18:27:03.249304512Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/stages/docx_extractor_stage.py", line 17, in <module>
2025-03-31T18:27:03.249309377Z     from nv_ingest.extraction_workflows import docx
2025-03-31T18:27:03.249311460Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/extraction_workflows/docx/__init__.py", line 6, in <module>
2025-03-31T18:27:03.249313689Z     from nv_ingest.extraction_workflows.docx.docx_helper import python_docx
2025-03-31T18:27:03.249315665Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/extraction_workflows/docx/docx_helper.py", line 31, in <module>
2025-03-31T18:27:03.249317822Z     from nv_ingest.extraction_workflows.docx.docxreader import DocxReader
2025-03-31T18:27:03.249319761Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/extraction_workflows/docx/docxreader.py", line 50, in <module>
2025-03-31T18:27:03.249321923Z     from nv_ingest.extraction_workflows.image.image_handlers import extract_page_elements_from_images
2025-03-31T18:27:03.249324056Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/extraction_workflows/image/__init__.py", line 1, in <module>
2025-03-31T18:27:03.249326312Z     from .image_handlers import image_data_extractor as image
2025-03-31T18:27:03.249328264Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/extraction_workflows/image/image_handlers.py", line 34, in <module>
2025-03-31T18:27:03.249330403Z     import nv_ingest.util.nim.yolox as yolox_utils
2025-03-31T18:27:03.249333297Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/nv_ingest/util/nim/yolox.py", line 17, in <module>
2025-03-31T18:27:03.249335601Z     import cv2
2025-03-31T18:27:03.249337614Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/cv2/__init__.py", line 181, in <module>
2025-03-31T18:27:03.249339991Z     bootstrap()
2025-03-31T18:27:03.249342755Z   File "/opt/conda/envs/nvingest/lib/python3.10/site-packages/cv2/__init__.py", line 153, in bootstrap
2025-03-31T18:27:03.249348820Z     native_module = importlib.import_module("cv2")
2025-03-31T18:27:03.249351226Z   File "/opt/conda/envs/nvingest/lib/python3.10/importlib/__init__.py", line 126, in import_module
2025-03-31T18:27:03.249353429Z     return _bootstrap._gcd_import(name[level:], package, level)
2025-03-31T18:27:03.249355405Z ImportError: libGL.so.1: cannot open shared object file: No such file or directory
2025-03-31T18:27:03.249357433Z 
2025-03-31T18:27:03.249359591Z ERROR conda.cli.main_run:execute(125): `conda run python main.py` failed. (See above for error)

IMO the project would greatly benefit from a docker-composeable minimal example that bootstraps the client library. Like a sample/hello-ingest.py type deal. Or better yet, a sandbox in browser that lets me upload a couple of my test docs and e2e see the results. My only goal here is to evaluate how well nv-ingest does on some of our existing test docs.

Thanks

[melo-gonzo]

@hughesadam87 You may need to install some additional base packages to be able to use open-cv in your container.
RUN apt-get update && apt-get install ffmpeg libsm6 libxext6 -y

[hughesadam87]

Thanks.

For what it's worth, I also just realized that some nv-ingest client libraries are on PyPI and not exclusively available through conda.

• nv-ingest-api: https://pypi.org/project/nv-ingest-api/
• nv-ingest-client: https://pypi.org/project/nv-ingest-client/

But not the library nv-ingest. Am I missing something? How can I get this line of code to work from pip installs

from nv_ingest.util.pipeline.pipeline_runners import start_pipeline_subprocess

This is the dockerfile I have so far without conda, but it's failing because pip can't find nv_ingest

FROM python:3.10-slim

WORKDIR /app

RUN apt-get update && apt-get install -y --no-install-recommends gcc ffmpeg libsm6 libxext6 && \
    rm -rf /var/lib/apt/lists/*

# Upgrade pip and install all required packages from PyPI
RUN pip install --upgrade pip && \
    pip install nv-ingest-api \
                nv-ingest-client \
                opencv-python \
                llama-index-embeddings-nvidia \
                pymilvus \
                "pymilvus[bulk_writer,model]" \
                milvus-lite \
                nvidia-riva-client \
                unstructured-client

# Copy the application code into the container
COPY . /app

LIke I said, many orgs have issues w/ conda these days and having a vanilla docker build of the client w/ a few hello would sample docs would be SO helpful. The notebooks are nice, but users like myself w/ windows and a company laptop are going to struggle to get a local env. to use them.