Point of the talk: To show how to upgrade traditional string-matching enrichment pipelines by injecting unstructured contextual data using LLM structured outputs and vector databases.

Duration: 45 Minutes

Detailed Breakdown: Traditional enrichment relies on exact key matching (e.g., matching IDs or exact words). Modern pipelines require semantic enrichment—categorizing customer feedback, sentiment-tagging tickets, or merging disparate data based on intent. This talk outlines a production-ready architecture using Pydantic (for strict schema validation of LLM outputs) and vector databases for fast similarity lookups. We will discuss how to minimize LLM token costs through semantic caching and how to handle schema data drift safely.