What's changed: Added per-section figures (cert-figure-retrofit). New AI-103 Chapter 6 (Azure Content Understanding = multimodal structured extraction/analyzer & schema/prebuilt vs custom/build flow, generative-prompt extraction of entities/topics/summaries = structured output/grounding & evaluation/structured vs schema extraction)
6.2Extracting Entities, Topics, and Summaries with Generative Prompting
Understand extracting entities, topics, and summaries flexibly via prompts to a generative model—receiving structured output, grounding, evaluation, and when to use this vs prebuilt/custom.
Beyond Content Understanding, AI-103’s information-extraction domain also covers flexible extraction via generative prompting. Instruct a generative model to "extract entities, main topics, and a summary from this document," and receive it as structured output (a JSON schema, from Chapter 2) so the app can consume extraction not bound to predefined categories. It excels on long or unstructured content but risks hallucination, so pair it with grounding (base it on the source document) and evaluation.
When to use which: (1) structured, repetitive fields → Content Understanding (schema extraction) or prebuilt/custom. (2) unstructured, flexible, summarization or semantic extraction → generative prompting + structured output. (3) Choose by accuracy/cost/consistency needs. In practice you often combine them (extract → generate a summary, etc.).
Common: (1) "flexibly generate-extract a summary/main topics from a document" = generative prompting (receive as structured output). (2) Curb errors with grounding + evaluation. (3) Repetitive structured fields suit Content Understanding (schema). (4) Extraction (pull out) and generation (write a summary) combine but are distinct roles.
Watch out: (1) choose generative-prompt extraction (flexible/unstructured) vs schema extraction (structured/consistent) by requirements—neither is always right. (2) Generative extraction risks hallucination (use grounding, citations, evaluation). (3) Without specifying structured output, parsing results becomes unstable.
6.2.1Section summary
- Use generative prompting to flexibly extract entities/topics/summaries, received as structured output
- Curb hallucination in generative extraction with grounding + evaluation
- Structured = Content Understanding (schema); unstructured/summary = generative prompting; combine in practice
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which suits flexibly extracting summaries/main topics from long unstructured text without predefined categories?
Q2. What should you pair with generative-prompt extraction to curb plausible errors (hallucination)?
Q3. Which suits consistently and repeatedly extracting structured fields from many similar documents?
Q4. What should you specify to reliably parse and use a generative model’s extraction results in an app?
Keep track of your progress
The full study guide is free to read. Sign up free to practice with the question bank, track what you have read, review your mistakes, and highlight passages.