Azure AI Search Knowledge Store Projections: JSON vs Extracted Text

Unstructured JSON requires Object Projection The first data source is unstructured JSON. Azure AI Search uses Object projection to store complex, hierarchical JSON documents directly into the knowledge store as individual blob files. This preserves the original structure and allows downstream applications to consume the raw JSON without parsing it from a table row.

Extracted Text from Scanned PDFs requires Table Projection The second data source is text extracted from scanned PDF documents. While the input was binary, the enrichment pipeline has already performed OCR (Optical Character Recognition) to extract text. The resulting text is typically stored in a structured format. In the context of Azure AI Search knowledge stores, extracting content like text or key-value pairs from documents is best handled by Table projection, which stores the extracted data in relational tables for efficient querying and indexing.

Why File Projection is Incorrect File projection is specifically designed for storing normalized, binary image files (e.g., resizing images or extracting thumbnails). It is not used for JSON data or extracted text strings. Using file projection for text would be inefficient and incorrect because text is not a binary image file.

Addressing Common Confusion Some learners might consider "Object projection" for the text if they view it as just another string payload. However, since the text is extracted content intended for search indexing, storing it in a table via Table projection aligns with how Azure AI Search indexes and queries structured data. The JSON, being inherently hierarchical and unstructured, remains an object.