← AI-103: AI apps and agents on Azure
06 / 6 · 40 MIN

Ingestion, retrieval, and extraction

Build useful, traceable representations for agents and reconciliation.

Concept and mechanism

A grounding chain begins at the source and ends at retrieved content. Scanned documents can need OCR; layout helps retain reading order, tables, and relationships. Chunking should retain useful units such as condition and action within context limits. Evaluate size and overlap using representative queries. Text search favors exact terms; vectors can help retrieve semantically similar descriptions. Hybrid combinations should be compared against actual query patterns. Reranking acts on existing candidates and does not replace correct ingestion. Vector-compression changes require measuring relevance, recall, and latency alongside space reduction. More results or HTTP 200 do not establish quality.

Guided application

In fictional reconciliation, define typed fields, currency, units, missing-value handling, and validation rules. Content Understanding can produce structured representations or content suitable for downstream reasoning according to analyzer and version. Configure supported detail and source options when reviewers need to locate values. Compare known documents and retain the link between field, document, and exception. High confidence does not authorize the operation or remove review of business criteria. Trace a document from OCR and enrichment through the index to the answer. Also check updates and deletions: a revised source does not guarantee every derived representation already reflects the change. Hand APS counts, quality indicators, and owners for each stage.

IN PRACTICE

The index returns the right document, but the value lost its currency during transformation. Correction should target the stage where that information disappeared.

Common pitfalls

Visible text as an extracted text layer; compression as benefit without tradeoffs; confidence as traceability; updated source as every copy updated.

Related topics: Architecture, project, and release · Generative applications and evaluation · Agents, memory, and approvals

Take this idea with you

Link every value and answer to a verifiable source, version, and transformation.

Create account

Reference: Analyzer schema workflow output and evidence configuration · AI-103 skills measured2026-04-16; Microsoft Foundry documentation2026-09-30