{"id":16214,"date":"2025-05-23T01:23:18","date_gmt":"2025-05-23T08:23:18","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/ise\/?p=16214"},"modified":"2025-05-23T01:32:43","modified_gmt":"2025-05-23T08:32:43","slug":"durable-functions-for-rag-indexing","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/ise\/durable-functions-for-rag-indexing\/","title":{"rendered":"Durable Functions for Indexing in RAG: A Practical Python Approach"},"content":{"rendered":"

Durable Functions for Indexing in RAG: A Practical Python Approach<\/h1>\n

Have you ever tried building an indexing pipeline for a Retrieval-Augmented\nGeneration<\/a> (RAG) app and struggled to\nchoose between \u201cpush\u201d and \u201cpull\u201d<\/a>? Both have their\nadvantages, but each comes with its own challenges. In this article, we\u2019ll introduce a hybrid approach that combines the best of\nboth while keeping overhead low\u2014using Azure Durable\nFunctions<\/a> (in Python) to get the\nright balance. We’ll have a look at the strengths and weaknesses of existing methods before diving into how Durable Functions can\naddress common issues like scaling, state management, and retries\u2014without adding unnecessary complexity.<\/p>\n

Setup, Deployment, and Prerequisites<\/h2>\n

Before you dive in, note that all setup details are provided in our sample\nrepository<\/a>. We use azd<\/a><\/strong> and Bicep<\/a><\/strong> to set up the entire\ninfrastructure\u2014including AI Search<\/a>, Azure\nDocument Intelligence<\/a>,\nOpenAI embeddings, and more. A dev container<\/a> is provided to simplify the setup steps significantly, so you don\u2019t need to worry about specific package versions or\nmanual configurations.<\/p>\n

\n

Why Indexing Matters for RAG<\/h2>\n

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by grounding responses in external enterprise\ndata. The accuracy of the generated output depends entirely on the quality of the retrieved data\u2014garbage in, garbage out.<\/strong>\nWhile prompt engineering can fine-tune model behavior, it cannot compensate for irrelevant, outdated, or poorly structured data\nin the index.<\/p>\n

That\u2019s why a robust indexing pipeline<\/strong> is critical. Instead of relying solely on prompt tuning, focus on ensuring the\nright data is retrieved at the right time<\/strong>. A well-structured RAG pipeline ensures that:<\/p>\n

    \n
  1. Enterprise documents<\/strong> (like PDFs or reports) are ingested properly. <\/li>\n
  2. Content is extracted<\/strong> cleanly and structured for retrieval. <\/li>\n
  3. Chunks are meaningful and contextually relevant.<\/strong> <\/li>\n
  4. Embeddings are high-quality, ensuring accurate retrieval.<\/strong> <\/li>\n
  5. The index is continuously refined<\/strong> to reflect the latest ground truth. <\/li>\n<\/ol>\n

    This is where Azure Durable Functions<\/strong> come into play. By automating document ingestion, processing, and indexing, you ensure\nthe most relevant and up-to-date<\/strong> information is always available for retrieval. This leads to more reliable, fact-based\nresponses from the LLM, reducing the risk of hallucinations or outdated results.<\/p>\n

    By investing in indexing, rather than overfitting prompts to imperfect data, you create a scalable, adaptable<\/strong> RAG system\nthat remains robust as your enterprise data evolves.<\/p>\n

    \n

    Push vs. Pull: The Usual Approaches<\/h2>\n

    Push Method<\/h3>\n