How does it connect different wording?

A question about “customers leaving” can match a document about “churn.” Meaning based representations help when users do not know the document's wording. They are useful for prose, support questions, and discovery.

How are candidates selected?

The system encodes documents and the query, searches for similar representations, and may rerank candidates using the full query and passage. Metadata filters can limit dates, regions, document types, or permissions. Chunking and model choice affect results. MIP's architecture overview shows why retrieval and business meaning work together.

Why is semantic similarity insufficient for metrics?

Two passages may discuss sales while using different definitions of revenue. Search locates candidates; a governed semantic model decides which metric and calculation apply. Similar wording does not prove identical business meaning. Hybrid search combines meaning based results with exact matches when both signals matter.

FAQ

Will it understand internal terms?

Not always; add term mappings or domain definitions.

Can it calculate exact numbers?

It can find relevant material, but calculation belongs in a database query.

How is it evaluated?

Test recall, ranking, and false matches on real questions.