Skip to content
Mitra Integrasi Informatika

AI Apps

Azure AI Search

The retrieval layer under a knowledge assistant, combining keyword and vector search with the access rules that decide what a given user may be answered from.

Product

Azure AI Search

Azure AI Search is the retrieval layer a grounded assistant reads from. It indexes documents and records as searchable fields, supports keyword search, vector search, and the hybrid of the two, and applies a semantic ranker over the result set so the passages that actually answer the question rise above the ones that merely share its words. Security trimming applies the source document's own entitlement at query time, so a user is never answered from a document they may not open. Because the retrieved passages are returned rather than summarised away, the assistant can cite what it used.

How it works

  • Documents are extracted and split into passages sized for retrieval, with the source location and the entitlement carried on each one.
  • Each passage is indexed as text and as a vector, so a question can be matched on wording and on meaning.
  • A query runs keyword and vector search together, and the hybrid result set is reranked semantically against the question that was asked.
  • Security trimming filters the result set by the user's identity, so entitlement is enforced at query time and not by trusting the prompt.
  • The selected passages are returned to the application, which assembles the grounded prompt and keeps the citations alongside it.
  • The answer is shown with its sources, so a reader can open the document rather than take the assistant's word for it.
  • Unanswered and low confidence questions are reviewed, and they become the backlog for the next indexing or chunking change.

A problem it solves

A knowledge assistant answers confidently from whatever text it happened to retrieve, and nobody can tell whether the answer came from the current policy or a superseded draft, or whether the person asking was entitled to that document at all. Azure AI Search makes retrieval the controlled step: what was searched, how it was ranked, whose entitlement applied, and which passages the answer was built from.

Where your problem lands

Have an enterprise challenge to solve?

Start with the problem. We will connect you to the practice accountable for that outcome, with the context you have already given.