docs: add Azure AI Search + Azure OpenAI RAG recipe notebook #675
Conversation
Merge ProtectionsYour pull request matches the following merge protections and will not be merged until they are valid. 🟢 Enforce conventional commitWonderful, this rule succeeded.Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/
|
|
@vagenas @pablocastro can you review :) |
|
@vagenas plz merge when ready |
| " raise EnvironmentError(\n", | ||
| " \"No GPU or MPS device found. Proceed with CPU only if you understand the performance implications.\"\n", | ||
| " )" |
There was a problem hiding this comment.
I see the intent here, but raising an error in case of no GPU could be a bit too much — could perhaps just print the message instead.
| "source": [ | ||
| "from openai import AzureOpenAI\n", | ||
| "from azure.search.documents import SearchClient\n", | ||
| "import uuid\n", |
There was a problem hiding this comment.
uuid is not used, I would drop this line.
| "# Part 2: Parse the PDF with Docling\n", | ||
| "Example: \"State of AI\" slides from a remote link.\n", | ||
| "\n", | ||
| "You can find the raw powerpoint here: https://docs.google.com/presentation/d/1GmZmoWOa2O92BPrncRcTKa15xvQGhq7g4I4hJSNlC0M/edit?usp=sharing" |
There was a problem hiding this comment.
I would use the same URL that is actually parsed below (which is a PDF and not a PowerPoint).
| "\n", | ||
| "console = Console()\n", | ||
| "\n", | ||
| "source_url = \"https://ignite2024demo.blob.core.windows.net/state-of-ai-2024/State of AI Report 2024.pdf\"\n", |
There was a problem hiding this comment.
- I see this is 212 pages of PDF. If the size was explicitly meant to be in that range, it could help to include a quick headsup to the user explaining that and that they would be parsing 200+ pages of PDF. From my point of view a shorter document could would work here.
- If we stick to this document, I would still URL-encode the URL.
| "outputs": [], | ||
| "source": [ | ||
| "# If running in a new environment, uncomment and run these:\n", | ||
| "%pip install docling~=\"2.7.0\"\n", |
There was a problem hiding this comment.
I see further below that HierarchicalChunker is imported by docling_core import package, but the docling-core distribution package is not installed here. For simplicity, I would:
- require a minimum of docling 2.9.0 (i.e.
docling~="2.9"), and - then import
HierarchicalChunkerasfrom docling.chunking import HierarchicalChunker
~= notation (e.g. docling~="2.9.0") it will only consider that minor version (2.9) and not include e.g. 2.10 etc.
| "%pip install azure-search-documents==11.5.2\n", | ||
| "%pip install azure-identity\n", | ||
| "%pip install openai\n", | ||
| "%pip install rich\n", | ||
| "%pip install torch" |
There was a problem hiding this comment.
I would collapse all install statements into a single one.
| } | ||
| ], | ||
| "source": [ | ||
| "from docling_core.transforms.chunker import HierarchicalChunker\n", |
There was a problem hiding this comment.
The import line discussed further above.
| }, | ||
| "source": [ | ||
| "# RAG with Azure AI Search\n", | ||
| "[](https://colab.research.google.com/github/farzad528/docling/blob/main/docs/examples/rag_azuresearch.ipynb)\n", |
There was a problem hiding this comment.
Let's change the URL to the target one that should work after merging, i.e. using DS4SD instead of farzad528
| "GPU or MPS usage can speed up Docling’s parsing (especially for large PDFs or when OCR/table extraction is needed). However, if no GPU is detected, you can comment out the following checks and proceed with CPU, albeit slower performance." | ||
| ] | ||
| }, | ||
| { | ||
| "cell_type": "code", | ||
| "execution_count": 4, | ||
| "metadata": {}, | ||
| "outputs": [ | ||
| { | ||
| "name": "stdout", | ||
| "output_type": "stream", | ||
| "text": [ | ||
| "CUDA GPU is enabled: NVIDIA A100 80GB PCIe\n" | ||
| ] | ||
| } | ||
| ], | ||
| "source": [ | ||
| "import torch\n", | ||
| "\n", | ||
| "if torch.cuda.is_available():\n", | ||
| " device = torch.device(\"cuda\")\n", | ||
| " print(f\"CUDA GPU is enabled: {torch.cuda.get_device_name(0)}\")\n", | ||
| "elif torch.backends.mps.is_available():\n", | ||
| " device = torch.device(\"mps\")\n", | ||
| " print(\"MPS GPU is enabled.\")\n", | ||
| "else:\n", | ||
| " # Comment out the error if you'd like to allow CPU fallback\n", | ||
| " # But be aware parsing could be slower\n", | ||
| " raise EnvironmentError(\n", | ||
| " \"No GPU or MPS device found. Proceed with CPU only if you understand the performance implications.\"\n", | ||
| " )" | ||
| ] |
There was a problem hiding this comment.
Starting with version 2.12, Docling provides more extensive accelerator support — and automatically detects and enables any available accelerator device.
You could therefore require ~="2.12" in the pip installs above and replace this section with some simple heads-up / disclaimer to the users.
| "AZURE_SEARCH_ENDPOINT = os.getenv(\"AZURE_SEARCH_ENDPOINT\")\n", | ||
| "AZURE_SEARCH_KEY = os.getenv(\"AZURE_SEARCH_KEY\")\n", | ||
| "AZURE_SEARCH_INDEX_NAME = os.getenv(\"AZURE_SEARCH_INDEX_NAME\")\n", | ||
| "AZURE_OPENAI_ENDPOINT = os.getenv(\"AZURE_OPENAI_ENDPOINT\")\n", | ||
| "AZURE_OPENAI_API_KEY = os.getenv(\"AZURE_OPENAI_API_KEY\") # Ensure this your Admin Key\n", | ||
| "AZURE_OPENAI_CHAT_MODEL = os.getenv(\"AZURE_OPENAI_CHAT_MODEL\") # Using a deployed model named \"gpt-4o\"\n", | ||
| "AZURE_OPENAI_API_VERSION = os.getenv(\"AZURE_OPENAI_API_VERSION\", \"2024-10-21\")\n", | ||
| "AZURE_OPENAI_EMBEDDINGS = os.getenv(\"AZURE_OPENAI_EMBEDDINGS\") # Using a deployed model named \"text-embeddings-3-small\"" |
There was a problem hiding this comment.
This section with os.getenv() is not going to work on Colab. Check out how we address this e.g. here (see _get_env_from_colab_or_os().
|
@farzad528 thanks for this nice example! 🙌 In order to align and merge:
|
Azure AI Search RAG Example Using Docling
This notebook demonstrates how to build a Retrieval Augmented Generation (RAG) system using Docling and Azure AI Search. The example showcases document parsing, chunking, vector search integration, and RAG query implementation.
Checklist: