Enterprise AI Agents with Structured and Unstructured Data Integration

snowflake · Tech · year n/a · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise data agents

Snowflake's talk addresses the gap between RAG over unstructured documents and enterprise agents that also need structured databases, semantic models, ambiguity resolution, and query planning to answer operational business questions.

Constraints:

RAG over documents is necessary but insufficient for enterprise agents.
Enterprise data lives across hundreds or thousands of tables with many columns and machine-oriented names.
Ambiguous user language must be mapped to business semantics and physical database values.
Agents can generate poor SQL if overloaded with raw schema details.
Structured and unstructured context must be combined for operational recommendations.

Architecture

Snowflake describes an agentic workflow that retrieves unstructured context, identifies relevant structured-data needs, resolves business semantics, generates a logical/human-readable representation of the data model before SQL, queries systems of record, and combines structured results with documents or Slack conversations for business recommendations.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  user-or-agent-request["User or agent request"] --> unstructured-retrieval["Unstructured retrieval"]
  unstructured-retrieval --> semantic-model["Semantic model"]
  semantic-model --> logical-data-model["Logical data model"]
  logical-data-model --> sql-query-generator["SQL/query generator"]
  sql-query-generator --> structured-databases["Structured databases"]
  structured-databases --> business-recommendation["Business recommendation"]
  unstructured-retrieval --> business-recommendation["Business recommendation"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`User or agent request`	Business question or task that needs enterprise context.	`Unstructured retrieval`
`Unstructured retrieval`	Searches documents, Slack conversations, SharePoint, PDFs, and other text/image stores.	`Semantic model`, `Business recommendation`
`Semantic model`	Maps user language to business concepts and database representations.	`Logical data model`
`Logical data model`	Transforms machine-readable schema into human-readable terms for query generation.	`SQL/query generator`
`SQL/query generator`	Creates queries against structured systems of record.	`Structured databases`
`Structured databases`	Enterprise systems of record containing sales, revenue, manufacturing, employee, and forecast data.	`Business recommendation`
`Business recommendation`	Combines structured and unstructured evidence into operational insight.	—

Data flow

A user or agent asks a business question; the workflow retrieves unstructured context, resolves terms such as country names through a semantic model, narrows the relevant structured data, generates a human-readable logical layer, creates SQL, queries systems of record, and combines results with documents or Slack context.

“in order for enterprises to reach full potential of these agents, you have to bring together the structured data with the unstructured data.”transcript, stated

“we actually flow it through this entire workflow of various different steps.”transcript, stated

Why this architecture

Snowflake argues that enterprise agents need both retrieval and structured-query planning because documents alone cannot answer questions about live sales, revenue, manufacturing, or forecast data.

“this is incredibly powerful. This is one of the essential building blocks that enterprises need for an agent to work. But it is not enough”transcript, stated

“I need to access those databases. I need to access those systems of record and query that data in order to get the right context back for an agent.”transcript, stated

Models used: LLMs · RAG · SQL generation · Snowflake Cortex

Scale

The talk says enterprises often have hundreds or thousands of tables.
“Now, enterprises often have hundreds if not thousands of different tables.”transcript, stated
The data models may reflect 20, 30, or 40 years of machine-readable enterprise database design.
“how enterprises have been building databases for the last 20, 30, 40 years is to be machine readable.”transcript, stated
The demo generated a query in a few seconds for sales, manufacturing, and forecast data.
“it generated the query in a few seconds so that I can start to access things like sales data and manufacturing data and forecast data.”transcript, stated

Alternatives considered (roads not taken)

Use RAG over unstructured documents alone.

Rejected because: Document retrieval cannot answer questions requiring structured systems of record.

“if I want to understand a question like what was revenue like in the last week, I can't just go search all my documents and use rag”transcript, stated

Send the raw structured-data request straight to the model.

Rejected because: Snowflake says the workflow should pass through several steps to provide the right context.

“we don't just send that straight to an underlying model and say give me the data.”transcript, stated

Ask enterprises to make old database schemas human-readable.

Rejected because: The talk calls that a herculean task and instead proposes a generated logical layer.

“I don't want to go ask all these enterprises, go change your data model and make it all human readable. That's a herculean task.”transcript, stated

Patterns

`rag-retrieval` (retrieval)

The workflow still uses retrieval over documents, Slack, SharePoint, PDFs, and other unstructured content.

“you take a tremendous amount of documents. Maybe these are Google Drive documents, Slack conversations, SharePoint, PDFs, you name it.”transcript, stated

`tool-use-orchestration` (orchestration)

The agent workflow orchestrates retrieval, semantic mapping, logical-schema generation, SQL, and data combination steps.

“we actually flow it through this entire workflow of various different steps.”transcript, stated

`query-rewriting` (retrieval)

The workflow rewrites physical database terms into a logical human-readable layer before SQL generation.

“go generate a human readable or a reasonable representation of the parts of the data model that you mean.”transcript, stated

Tradeoffs

Use multi-step agentic data access.

Chose Semantic and logical-model steps before database querying. over Direct model-to-database querying..

Rationale: The extra steps improve context quality and query generation against messy enterprise schemas.

“each of these different steps will make a little bit more sense.”transcript, implied

Generate a logical layer instead of changing databases.

Chose Agent-generated human-readable representation. over Rewriting enterprise data models..

Rationale: Reworking decades of database design would be too large a task for adoption.

“go change your data model and make it all human readable. That's a herculean task.”transcript, stated

Gotchas (what bit them)

Unstructured RAG alone misses structured business facts.

Root cause: Revenue, sales, employees, and operational facts are stored in databases rather than documents.

Resolution: Bring structured systems of record into the agentic workflow.

“There's revenue data. There's sales data. There's employee data. There's data about every facet of an organization”transcript, stated

Ambiguous business terms break direct querying.

Root cause: User language like 'US' may not match the database's representation of the same concept.

Resolution: Use a semantic model that maps user language to database values and concepts.

“Well, do I mean United States? And if I mean the United States, is my data, if I go look at the database, is that what the column has US called?”transcript, stated

Raw schemas overwhelm query generation.

Root cause: Enterprise schemas are often machine-readable and full of technical column names.

Resolution: Generate a logical layer with readable names and relationships before SQL generation.

“if you throw all these kind of technical jargon terms at an LLM and then say go create a SQL query, the quality is significantly less”transcript, stated

Evaluation

The talk evaluates the architecture through quality of generated SQL/query results and demo speed on a Whoop-style business question combining structured and unstructured data.

Metrics: Query generation quality · Seconds to generate a query · Ability to combine sales, manufacturing, forecast, and Slack context

“if you ask an agent now to generate some SQL based on these characteristics, it gives much higher quality results”transcript, stated

Outcomes

The demo agent could query sales, manufacturing, and forecast data in seconds.
The architecture combines structured and unstructured context for business recommendations.
The case argues enterprise agents need database-aware workflows, not just document RAG.

The enterprise setting includes hundreds or thousands of tables and decades of machine-readable database design.

“The agent can then reason around that. It can combine it with unstructured data.”transcript, stated