Unlock your Dataverse data with secure and powerful search | DEM523

by Unknown (Microsoft Dataverse team)

Day 3 · Demo Session · 16 min · 3 min read

Applied AI NLP & Language

TL;DR

Dataverse unified search, demonstrated live at Microsoft Build 2025, adds fuzzy matching and recency-weighted ranking to keyword search, then scopes results per app. A data exploration agent converts plain-English questions into SQL: one query filtered October online orders and rendered a pie chart where chocolate-filled doughnuts took 50% of revenue. A file-attachment query returned baking temperature from a recipe PDF. Dataverse search then moves to MCP, where Claude queried customer reviews and produced recipe improvement recommendations with no hand-written query.

The Search Problem in Business Data Platforms

▶ Watch (00:11)

Dataverse sits behind all of Microsoft’s Dynamics and Power Platform applications. Every sales record, customer service ticket, and Power Automate run lands there. At production scale that is hundreds of GB of data, millions of records, and trillions of activity events for a single organization. The challenge is giving every employee, from front-line staff to analysts, a way to query that data without writing code or understanding the underlying tables.

Fuzzy Match, Relevance Ranking, and App-Scoped Filtering

▶ Watch (02:36)

Three gaps show up in keyword search: typos return nothing, exact matches flood results when hundreds of records share one word, and two apps sharing a table return the same rows regardless of app scope. Dataverse unified search addresses each one. A one-character typo still finds the record. The ranking model weights recently touched rows higher. Per-app filters trim the scope, so an online-only app returns 3 matching orders while the main shop returns 5 for the same name.

Natural Language to SQL: The Data Exploration Agent

▶ Watch (05:01)

The data exploration agent converts plain-English questions into SQL. Type “show me online orders in October” and it applies a date-range filter with online-order status set to yes. The NL pie chart: chocolate filled = 50% of October online orders (06:19) shows the result rendered as a pie chart.

“It takes about nearly 50% of the contribution, right.” — Unknown (Microsoft Dataverse team)

The same agent extends to unstructured data. Attach recipe PDFs to product records and ask “what is the baking temperature,” and it retrieves the file and returns the extracted answer. No schema knowledge required.

Dataverse as an MCP Tool

▶ Watch (09:58)

Dataverse search is now exposed as a tool inside an MCP server. An LLM gets a list of available tools and decides which to call. You don’t write the query logic.

The MCP recipe improvement demo (11:03) shows Claude querying customer reviews from Dataverse with no hand-written query, then producing recipe recommendations from the summarized feedback. A second prompt asks Claude to compare online versus offline orders and recommend next steps.

“You just list out the tool to the MCP.” — Unknown (Microsoft Dataverse team)

Notable Quotes

Maple bacon doughnuts only sold like 6 units. Unknown (Microsoft Dataverse team) · ▶ 8:09

It takes about nearly 50% of the contribution, right. Unknown (Microsoft Dataverse team) · ▶ 6:45

You just list out the tool to the MCP. Unknown (Microsoft Dataverse team) · ▶ 13:29

Key Takeaways

Dataverse unified search adds fuzzy matching and recency-based ranking on top of standard keyword search.
The data exploration agent converts plain-English questions into SQL without requiring users to know the schema.
Exposing Dataverse as an MCP server lets LLMs query, analyze, and act on live business data autonomously.

Filed under

Applied AI NLP & Language