---
title: "Semantic layer for natural-language analytics"
url: "https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-for-natural-language-analytics/"
description: "How a natural language analytics semantic layer turns a question into one governed number, what six platforms read, and where Atlan's AI Context Platform fits."
author: "Emily Winks"
author_role: "Data Governance Expert"
published: "2026-11-17"
updated: "2026-11-17"
---

## What does a semantic layer do for natural-language analytics?

In natural-language analytics, people ask data questions in ordinary words and get a query result back. A semantic layer is the one shared place that defines each metric, its joins and its default filters, so the system answering the question uses the agreed calculation instead of guessing from raw tables. It cannot settle a question with two reasonable readings, or data that does not exist, so a well-built agent names the definition it used and asks when the question is unclear.

The four steps a question takes:

- **Resolve terms:** map the words to defined business terms
- **Apply the measure:** use one certified calculation
- **Add joins and filters:** the join path, grain and default filters
- **Show the route:** run the query and say which definition answered

## Key takeaways

- A semantic layer settles the measure, the joins and the default filters before a question becomes SQL.
- Snowflake, Databricks, Looker, Fabric, Power BI and dbt each offer a governed object an agent can answer from.
- Published studies measure different tasks, so read each result with its own setup.
- The layer cannot settle unforeseen ambiguity or missing data; the answer should name the definition it used.

---

For analytics and AI teams putting an agent in front of business users, a semantic layer gives each plain-language question one governed definition to land on. It works in four steps: resolve the business terms, apply one certified measure, add the joins and default filters, and show the route taken. According to Rumiantsau and Fokeev (2026), in [a preprint on a single dataset](https://arxiv.org/abs/2604.25149), a 4 KB markdown file of measures and rules lifted three models from 45.5 to 50.5% accuracy to 67.7 to 68.7%. In Atlan's AI Context Platform, that definition is written once in a context repository; Context Engineering Studio deploys it to Snowflake Cortex Analyst and Databricks metric views and renders it for dbt MetricFlow and Apache Ossie.

### Is Your Semantic Model Ready for AI Agents

Give it your semantic model and the questions your agents will ask. It returns a pass or gap verdict for SM-01 to SM-12 and the first fix. [Read the skill](/skills/semantic-model-readiness-check.md).

*Paste into a new chat*

```
Use the skill at https://atlan.com/skills/semantic-model-readiness-check.md to check whether our semantic model is ready for natural-language questions. Ask me for whatever it needs.
```

*Run once in a terminal*

```
curl -fsSL --create-dirs \
  -o ~/.agents/skills/semantic-model-readiness-check/SKILL.md \
  https://atlan.com/skills/semantic-model-readiness-check.md
```

*For an agent*

```
curl -fsSL https://atlan.com/skills/semantic-model-readiness-check.md
```


    Human
    Agent


  ClaudeChatGPTGeminiCursorOther
  Copy

---

Natural-language analytics means asking a data question in ordinary words and getting a query result back. The semantic layer sits underneath that question and supplies the definitions it lands on. Whether you need one is answered in [do you need a semantic layer](https://atlan.com/know/ai-agent/semantic-layer/do-you-need-a-semantic-layer/), and the either-or choice in [text-to-SQL vs semantic layer for self-serve analytics](https://atlan.com/know/ai-agent/semantic-layer/text-to-sql-vs-semantic-layer-for-self-serve-analytics/).

Two limits come first. The layer does not settle ambiguity nobody anticipated, and even with supplied definitions the one paired study got about a third of its answers wrong. For products, see the [best semantic layer tools](https://atlan.com/know/best-semantic-layer-tools/); for the basics, the [semantic layer guide](https://atlan.com/know/semantic-layer/).

| Field | Value |
|---|---|
| Persona | AI Leader, Analytics Leader |
| The job | Turn a plain-language question into one governed number |
| What the layer settles | Which measure, which joins, which default filters |
| What it leaves open | Which of two readings the user meant, and whether data exists for the question |
| Evidence used | Product docs from six platforms, and three published studies, each with its own setup |
| Status labels | Quoted only where a vendor's page prints one (Power BI Prep data for AI is Preview) |

---

## What does a semantic layer do when someone asks a question in plain language?

When a user types a question, the agent and the semantic layer work through four steps. They map the words to defined terms, pick the certified measure, add the joins and default filters, then run the query and record which definition answered.

Take "What was net revenue last quarter?" That one phrase hides two rules: a measure rule (what "net revenue" subtracts) and a time rule (fiscal or calendar quarter).

### One question, four steps

Atlan's documentation walks the same question through a context repository. The semantic model supplies the measure, joins and default filters. The linked glossary term supplies the certified definition, and custom instructions apply rules outside the model ([What's in a context repository, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/whats-in-a-context-repository)).

| Step | What happens | What goes wrong if this step is missing |
|---|---|---|
| 1. Resolve terms | "Net revenue" maps to a defined term and "last quarter" to a time rule | The model guesses which revenue column the user meant |
| 2. Apply the governed measure | One certified calculation for net revenue is used | Each tool writes its own formula, and two answers disagree |
| 3. Joins and default filters | The join path, the grain and filters such as excluding test accounts are added | A one-to-many join counts orders twice, or test data slips in |
| 4. Run and show the route | The query runs on the connected engine and the answer names the definition it used | The user cannot tell a certified answer from a guess |

![Flow showing a plain-language revenue question passing through four semantic layer steps to one governed number.](/images/semantic-layer-for-natural-language-analytics/1-one-question-through-a-semantic-layer.webp "One question through a semantic layer"){width=1536 height=1024}

One question through a semantic layer. Source: Atlan.

### Where the definition ends and the SQL begins

Snowflake's documentation describes the same division of labor for its own product: Cortex Agents "reads the information captured in the semantic view definition and generates the SQL against the physical tables directly" ([Overview of semantic views, Snowflake docs](https://docs.snowflake.com/en/user-guide/views-semantic/overview)). The definition is fixed; the SQL is still generated per question.

A [semantic layer for analytics](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-for-analytics/) does this for dashboards; here it serves a question nobody predicted. Whether you call it a [semantic layer or a metrics layer](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-vs-metrics-layer/), the [business glossary](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-vs-business-glossary/) holds the meaning and the [semantic model](https://atlan.com/know/ai-agent/semantic-layer/what-is-a-semantic-model/) holds the calculation.

### How is a semantic layer different from a presentation layer?

The presentation layer is what a person looks at: the dashboard, the report, the chart. The semantic layer sits beneath it and holds the metric definitions those views read. dbt's documentation describes moving metric definitions "out of the BI layer and into the modeling layer", so business units work from the same definitions whatever tool they use ([dbt Semantic Layer, dbt Developer Hub](https://docs.getdbt.com/docs/use-dbt-semantic-layer/dbt-sl)). An agent answering in chat skips the dashboard entirely, so a formula kept only in the presentation layer never reaches it.

### Why the same question gets two numbers

A dashboard and an agent often disagree because the agent picked its own measure or skipped a default filter. The dashboard runs a reviewed calculation. The agent built one on the fly. Rule out the plainer causes too, such as a different refresh time, row-level security that shows each user different rows, or a calendar quarter set against a fiscal one. The same pattern explains [why Genie and Copilot answers differ from BI reports](https://atlan.com/know/ai-agent/semantic-layer/why-genie-copilot-answers-differ-from-bi-reports/). It is why a [semantic layer serving both BI and AI agents](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-for-bi-vs-ai-agents/) needs one definition rather than two copies.

The context an agent reads beyond metrics is covered in [why natural-language analytics needs a context layer](https://atlan.com/know/ai-agent/context-layer/context-layer-for-natural-language-analytics/), and per-question readiness in [AI-ready data for talk-to-data](https://atlan.com/know/ai-agent/ai-ready-data/ai-ready-data-for-talk-to-data/).

---

## How do Snowflake, Databricks, Looker, Fabric, Power BI and dbt each handle the same question?

Each major platform offers a governed object a plain-language question can resolve through, names it differently, and adds example queries or instructions around it. Some, such as the Fabric data agent and Genie Agents, can also answer from tables and warehouses directly.

### What each platform's docs say

Rows quote only each vendor's docs, opened 2026-11-17.

| Platform and product | Object that resolves the question | What adds accuracy | Scope or status the docs state | Source |
|---|---|---|---|---|
| Snowflake Semantic Views, Cortex Agents | Semantic view: a schema-level object holding metrics, entities and relationships | Verified ("gold") queries, suggested verified queries from real usage, custom instructions | Used by Cortex Agents, or queried in a SELECT statement | [Snowflake docs](https://docs.snowflake.com/en/user-guide/views-semantic/improving-accuracy) |
| Databricks Unity Catalog metric views, Genie Agents | Metric view: define a metric once, query it from Genie Agents, SQL or dashboards | Agent metadata (synonyms, display names, formats); example SQL, instructions, knowledge store | Up to 50 tables, views or metric views per Genie Agent; metric view relationships are labeled Beta | Databricks docs (see Sources) |
| Google Looker Conversational Analytics | The Looker semantic modeling layer (Explores) | Verified queries added to the data agent's context | A data agent can query as many as five Explores; Google recommends validating all Gemini output | [Google Cloud docs](https://docs.cloud.google.com/looker/docs/conversational-analytics-overview) |
| Microsoft Fabric data agent | Lakehouses, warehouses, Power BI semantic models, KQL databases, ontologies | Instructions and example queries | Read-only; respects Microsoft Purview controls; generally available | [Microsoft Learn](https://learn.microsoft.com/en-us/fabric/data-science/concept-data-agent) |
| Microsoft Power BI Copilot, Prep data for AI | The Power BI semantic model | AI data schemas, verified answers, AI instructions, "Approved for Copilot" | Preview, per its own title | [Microsoft Learn](https://learn.microsoft.com/en-us/power-bi/create-reports/copilot-prepare-data-ai) |
| dbt Semantic Layer, dbt MCP server | MetricFlow metrics | AI tools connect through the dbt MCP server (Model Context Protocol, an open standard for AI tool calls), exposing the Semantic Layer, SQL and Discovery API | Semantic Layer on Starter, Enterprise and Enterprise+ plans | [dbt docs](https://docs.getdbt.com/docs/dbt-ai/about-mcp) |

### What the six platforms have in common

Three levers recur across most rows: a governed object, example or verified queries, and instructions. According to dbt's documentation, AI tools that connect through its MCP server get answers that "use your governed metrics instead of guessing at raw tables" ([dbt Semantic Layer, dbt Developer Hub](https://docs.getdbt.com/docs/use-dbt-semantic-layer/dbt-sl)). That only holds when the agent is pointed at the governed object. Scope the agent's sources to semantic views, metric views or models where the platform allows it, and label any answer built from base tables as uncertified. Genie Agents were formerly called Genie Spaces, so older material uses that name.

For one platform in depth, see [Snowflake semantic views](https://atlan.com/know/snowflake/snowflake-semantic-views/), [Databricks metric views](https://atlan.com/know/ai-agent/semantic-layer/how-to-create-databricks-metric-views/), the [context Databricks Genie requires](https://atlan.com/know/ai-agent/databricks/databricks-genie-context-requirements/), [Fabric data agents](https://atlan.com/know/microsoft-fabric/what-are-fabric-data-agents/), the [context Power BI conversational analytics needs](https://atlan.com/know/ai-agent/microsoft/power-bi-conversational-analytics-context/) and the [MCP server for dbt](https://atlan.com/know/mcp/mcp-server-for-dbt/).

When the same metric is authored separately in two of these tools, it lives in two places and can drift. dbt's Semantic Layer already serves one centrally defined metric to several downstream tools. Apache Ossie, formerly Open Semantic Interchange, is an incubating Apache project for exchanging semantic models between tools ([Apache Ossie, GitHub](https://github.com/apache/ossie)). [Open Semantic Interchange](https://atlan.com/know/ai-agent/semantic-layer/what-is-open-semantic-interchange/) covers its origins, and [warehouse-native vs universal semantic layers](https://atlan.com/know/ai-agent/semantic-layer/warehouse-native-vs-universal-semantic-layer/) the architecture choice.

---

## What results have studies reported on semantic layers and text-to-SQL accuracy?

Three published studies bear on this, and they measure different things: one paired test of supplied business definitions, and two benchmarks of how hard enterprise text-to-SQL is without them. Their datasets, tasks and scoring differ, so each result stands with its own setup and none is a ceiling for your data.

| Study | What it measured | Setup | What the authors report |
|---|---|---|---|
| [Rumiantsau and Fokeev, arXiv 2604.25149](https://arxiv.org/abs/2604.25149) (preprint, April 2026) | Accuracy and hallucination (confident answers the data does not support), with and without business semantics, three frontier models | 100 questions on the Cleaned Contoso Retail Dataset in ClickHouse; the "semantic layer" is a 4 KB hand-authored markdown document of measures, conventions and disambiguation rules, not a deployed product | With the document, 67.7 to 68.7%; with the schema alone, 45.5 to 50.5% |
| [BEAVER, arXiv 2409.02038](https://arxiv.org/abs/2409.02038) (revised May 2026) | Text-to-SQL on private enterprise warehouses | 9,128 question-SQL pairs from real query logs, 812 tables, 19 domains | Low accuracy even for agentic frameworks; oracle hints helped (figures below) |
| [BIRD-History, Findings of EMNLP 2026](https://aclanthology.org/2026.findings-emnlp.593/) | Grounding underspecified questions with historical SQL scripts | 1,393 tasks across 11 databases | Systems struggle when a question relies on business logic and data conventions the schema does not capture |

### Why enterprise text-to-SQL stays hard

According to the [BEAVER authors (2026)](https://arxiv.org/abs/2409.02038), agentic frameworks with GPT-5.2 reached 10.8%, and 30.1% with all subtask annotations given as oracle hints. In their words, "a major bottleneck lies in correctly resolving these subtasks."

The [BIRD-History authors (2026)](https://aclanthology.org/2026.findings-emnlp.593/) point at the gap a semantic layer is built to fill: business logic the schema never states. Neither tested a semantic layer; they show why the missing business meaning matters.

The studies describe the kind of problem; your own questions measure the size of the gain. More on supplying business meaning: [text-to-SQL with AI](https://atlan.com/know/ai-agent/data-for-ai/text-to-sql-with-ai/), [text-to-SQL for the enterprise](https://atlan.com/know/ai-agent/data-for-ai/text-to-sql-for-enterprise/) and [how a context layer controls hallucination](https://atlan.com/know/ai-agent/context-layer/how-context-layer-controls-hallucination/).

---

## What can a semantic layer not settle in natural-language analytics?

A semantic layer settles which measure, which joins and which default filters apply, and it can encode known ambiguities through synonyms, instructions and verified queries. It cannot settle an ambiguity nobody anticipated, or whether the data to answer exists.

Wrong grain, join fan-out and a missed filter are the model-side errors the layer is built to prevent. An ambiguous metric needs disambiguation rules, which is what the 4 KB document in the Rumiantsau and Fokeev preprint held.

### What the layer settles and what it leaves open

Google's own Looker documentation says Gemini "can generate output that seems plausible but is factually incorrect" and recommends validating all output ([Conversational Analytics in Looker overview, Google Cloud](https://docs.cloud.google.com/looker/docs/conversational-analytics-overview)).

| Question type | Settled by the layer | Left open | What the agent should do |
|---|---|---|---|
| Metric definition | Yes, through one certified measure | Whether this metric is the one the question asked for | Name the metric in the answer |
| Join path | Yes, through defined relationships and grain | A question that spans models with no defined join | Say the join is undefined |
| Default filter | Yes, through named filters | A filter the user wanted but did not state | Apply the default and show it |
| Filter values and relative dates | Partly, through synonyms and a stated time rule | Which member "EMEA" means, and what date "last quarter" is counted from | State the resolved values in the answer |
| Two certified measures fit | Partly: both definitions are correct | Which one this user meant, such as bookings before or after returns | Name the measure in the answer, or ask |
| Unforeseen ambiguous wording | No | Which of two readings the user meant | Ask one clarifying question |
| Missing data | No | Whether the data exists at all | Say so; never invent a number |

### What the agent should say when it falls back

When no certified metric fits, the fallback runs in order: certified metrics first, AI-ready data for the rest, and a plain statement of which route was used.

Austin Kronz of Atlan: "you should always start with certified metrics where you have them." The agent then tells the user, in his words, "Here's which route I took." Keep the route short. For example: "Net revenue (certified), fiscal Q3, test accounts excluded." Build it from the query that actually ran. The model's own account of what it did is not proof.

The second route needs [AI-ready data](https://atlan.com/know/ai-agent/data-for-ai/what-makes-data-ai-ready/), which differs from [analytics-ready data](https://atlan.com/know/ai-agent/ai-ready-data/ai-ready-data-vs-analytics-ready-data/), and a way for it to [reach agents over MCP](https://atlan.com/know/ai-agent/ai-ready-data/ai-ready-data-mcp/). Where a number must be exact every time, send the question to a reviewed, tested report. That answers whether the [semantic layer is dead](https://atlan.com/know/ai-agent/semantic-layer/is-the-semantic-layer-dead/): it catches measure, join and filter errors, and a wrong reading of the question still needs a person or a clarifying question.

Atlan's short video on where AI analysts and dashboards each fit:





---

## How do you tell whether a semantic model is ready for natural-language questions?

A semantic model is ready when the questions your users actually ask resolve to one defined measure, one join path and one owner, which a check against twelve items shows before launch. Four actions get you there:

1. **Collect the twenty questions people ask most.** Pull them from BI usage, support channels and analyst requests, and agree the right answer for each with the metric's owner. Count an answer correct only when its result matches that agreed answer, and keep these questions apart from the examples the agent is shown.
2. **Run the readiness check on the model that should answer them.** The semantic model readiness check scores twelve items, SM-01 to SM-12. Eight block launch: agent scope, tested keys, declared joins, formulas with exclusions, explicit fiscal or calendar time, a named owner, tests against real questions, and access enforced at query time.
3. **Fix blockers one at a time.** Start with the gap that can return a wrong number, fix it, rerun the twenty questions, then move to the next. Launch only when no blocker is left, and rerun the questions on every model change.
4. **Scope one use case and test before widening.** A model for one team's questions is easier to test and correct than one model for the whole company. Widen only after the first scope holds.

### Two pitfalls to avoid

Two pitfalls come up again and again. Instructions that restate a metric drift away from the model, so keep the formula in the model and save instructions for rules it cannot hold. And a definition with no owner has nobody to correct it when an answer is wrong, so name one owner per measure.

A model that passes the check has its structure in place; the twenty questions prove the numbers. Preparation is covered in [how to prepare data for natural-language analytics](https://atlan.com/know/ai-readiness/how-to-prepare-data-for-natural-language-analytics/), the architecture in the [talk-to-data agent blueprint](https://atlan.com/know/ai-agent/talk-to-data-agent-blueprint/), and the [business context for AI](https://atlan.com/know/business-context-for-ai/) around the model in [context engineering vs semantic layer](https://atlan.com/know/ai-agent/semantic-layer/context-engineering-vs-semantic-layer/).

---

## How Atlan approaches the semantic layer for natural-language analytics

Atlan's AI Context Platform holds the semantic model in a context repository scoped to one use case, and generates each engine's version from that one model.

According to Atlan's documentation, the repository links to governed Atlan assets. On top, it adds a semantic model of tables with dimensions, time dimensions and facts; relationships; derived metrics built on facts and glossary terms; named filters; verified queries; and custom instructions ([What's in a context repository, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/whats-in-a-context-repository)).

- Each repository serves one use case. In the docs' own example, "Answer pipeline questions for the sales ops team" is a use case and "Sales" is not ([What's in a context repository, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/whats-in-a-context-repository)). A narrow repository is easier to test and to correct when a question fails.
- Drift is visible: an Activity tab lists upstream changes since the model was built, and Regenerate rebuilds it from the current state of the linked assets.
- Context Engineering Studio deploys to Snowflake Cortex Analyst and to Databricks metric views with a Genie space, and renders files for dbt MetricFlow and Apache Ossie, all from the same canonical model. For Snowflake, Preview YAML shows the generated file before you ship it ([Portability targets, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/references/portability-targets)).
- Some engines cannot represent a metric with a conditional filter. Atlan holds that metric back from the engine, reports which one and why, and keeps it in the canonical model ([Metrics that aren't emitted, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/references/not-emitted-metrics)).

### Reaching agents outside the engines

Agents that read the repository over MCP get the same definition without a deploy ([What is Context Engineering Studio?, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/what-is-context-engineering-studio)); see [Atlan MCP](https://atlan.com/know/what-is-atlan-mcp/). The wider [enterprise context layer](https://atlan.com/know/context-layer-enterprise-ai/) and the neighbors in [semantic layer vs data catalog](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-vs-data-catalog/) and [semantic layer vs knowledge graph](https://atlan.com/know/ai-agent/semantic-layer/semantic-layer-vs-knowledge-graph/) (a graph of business entities and how they relate) cover the rest. Your team writes net revenue once, and the build report shows which engines could not take which metrics.

[See what a context repository holds, and how an agent uses it to answer "What was net revenue last quarter?"](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/whats-in-a-context-repository)

---

## FAQs about semantic layers for natural-language analytics

### 1. What is a semantic layer in natural-language analytics?

A semantic layer is the shared definition of each metric, its joins and its default filters that a natural-language analytics tool reads before writing a query. It turns the words in a question into one agreed calculation, so the answer matches the number the business already reports.

### 2. Why does an AI analytics agent give a different revenue number than my dashboard?

Often the agent chose its own measure or missed a default filter the dashboard applies. A dashboard runs a reviewed calculation, while an agent without a certified metric assembles one from raw tables. Check refresh times, row-level security and fiscal versus calendar quarters too. Pointing the agent at the certified metric first, and showing which one it used, removes the most common cause.

### 3. What does an analytics agent do when no metric matches the question?

A well-built agent says which route it took, names the nearest certified metric, and asks one clarifying question. It should not return a number from a formula it assembled on the spot. The visible route lets the user correct the answer instead of trusting a guess.

### 4. Can a semantic layer make natural-language answers fully accurate?

No. In a 2026 preprint on a single dataset, a 4 KB document of measures and rules raised accuracy for three models, and about a third of answers were still wrong. The layer narrows metric and join errors, but ambiguous wording and missing data still need a clarifying question or a reviewed report.

### 5. Which questions should a semantic layer not be asked to answer?

Questions with two reasonable readings that nobody has written a rule for, questions about data that does not exist, and figures that must be exact every time but have no tested definition. The first needs a clarifying question, the second needs a plain "no data" answer, and the third belongs on a reviewed, tested report.

### 6. Where does a semantic layer sit in the data stack?

It sits between the stored tables and the tools or agents that ask questions of them. In Snowflake, a semantic view is a schema-level object inside the database. In dbt, metric definitions live in the modeling layer, the dbt project, rather than in each BI tool. Either way, every consumer reads the same definition.

### 7. Is there an open standard for sharing semantic models between tools?

Apache Ossie, formerly known as Open Semantic Interchange, is an incubating Apache project working on a vendor-neutral way to exchange semantic models between BI, AI and analytics tools. It is still incubating, so check which version a tool supports before relying on it for portability.

---

## Sources

1. [Semantic Layers for Reliable LLM-Powered Data Analytics: A Paired Benchmark of Accuracy and Hallucination Across Three Frontier Models, arXiv](https://arxiv.org/abs/2604.25149)
2. [BEAVER: An Enterprise Benchmark for Text-to-SQL, arXiv](https://arxiv.org/abs/2409.02038)
3. [BIRD-History, Findings of EMNLP 2026, ACL Anthology](https://aclanthology.org/2026.findings-emnlp.593/)
4. [Overview of semantic views, Snowflake](https://docs.snowflake.com/en/user-guide/views-semantic/overview)
5. [Improving accuracy with AI, Snowflake](https://docs.snowflake.com/en/user-guide/views-semantic/improving-accuracy)
6. [Cortex Analyst Verified Query Repository, Snowflake](https://docs.snowflake.com/en/user-guide/views-semantic/verified-query-repository)
7. Unity Catalog metric views, Databricks documentation (updated 2026-10-29)
8. Create and manage a Genie Agent, Databricks documentation (updated 2026-11-05)
9. [Conversational Analytics in Looker overview, Google Cloud](https://docs.cloud.google.com/looker/docs/conversational-analytics-overview)
10. [Fabric data agent concepts, Microsoft Learn](https://learn.microsoft.com/en-us/fabric/data-science/concept-data-agent)
11. [Prepare your data for AI to improve Copilot results (preview), Microsoft Learn](https://learn.microsoft.com/en-us/power-bi/create-reports/copilot-prepare-data-ai)
12. [dbt Semantic Layer, dbt Developer Hub](https://docs.getdbt.com/docs/use-dbt-semantic-layer/dbt-sl)
13. [About dbt Model Context Protocol (MCP) server, dbt Developer Hub](https://docs.getdbt.com/docs/dbt-ai/about-mcp)
14. [Apache Ossie (incubating), GitHub](https://github.com/apache/ossie)
15. [What's in a context repository, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/whats-in-a-context-repository)
16. [What is Context Engineering Studio?, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/what-is-context-engineering-studio)
17. [Portability targets, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/references/portability-targets)
18. [Metrics that aren't emitted, Atlan docs](https://docs.atlan.com/product/capabilities/governance/context-engineering-studio/references/not-emitted-metrics)

## Next steps

- To see Atlan's context layer in action first, watch a recorded demo: [Atlan context layer demos](https://atlan.com/context-layer-demo/)
- To see Atlan working on your own data stack, book a demo: [Book a demo with Atlan](https://atlan.com/forms/talk-to-sales-contact/)
