Context layer sources and learning
The context layer learns from four categories of input. Each source flows through the same learning pipeline regardless of format or origin.
Maintained context
Maintained context comes from systems that someone in your organization already owns and curates.
| Source | Description |
|---|---|
Data catalogs |
Business glossaries and catalogs from systems such as Atlan or Collibra. Definitions are adopted directly, not re-learned. |
Semantic views |
Semantic models defined in your data platform (for example, Snowflake, dbt, or Databricks) including metrics, entities, and join logic. Refer to Context layer and the semantic layer for how the semantic layer relates to the context layer. |
External context layers |
Context graphs or ontologies your organization already maintains, such as Databricks Ontology. ThoughtSpot consumes these rather than replacing them. |
User and profile systems |
Identity and role information from your HRIS, CRM, or directory. These provide stable profile facts such as team, region, and territory. |
Documents and artifacts |
Documents that capture business definitions, processes, and institutional knowledge. |
Built artifacts
Built artifacts are analytics objects that encode business decisions through their construction.
| Source | Description |
|---|---|
Liveboards and Answers |
Every saved formula and filter encodes a definition that someone stood behind. |
Alerts and KPIs |
Thresholds and monitors represent business rules: what matters, and when it matters. |
Prior BI estate |
Workbooks and dashboards from previous BI tools such as Tableau or Looker. These are rich in embedded definitions and can be migrated into ThoughtSpot to feed the context layer. |
Interactions
Interactions are traces of how people use data in practice.
| Source | Description |
|---|---|
Conversations |
What people ask the Spotter agents, how they rephrase questions, and what answers they accept. |
Query logs |
How data has actually been queried in the warehouse: the real joins, filters, and metric expressions in daily use. |
Product interactions |
Clicks, drill-downs, and other product usage patterns that reveal how people work with data. |
Ambient surfaces
Ambient surfaces are unstructured, continuously growing sources of organizational knowledge.
| Source | Description |
|---|---|
Messaging |
Discussions in Slack or Microsoft Teams about business processes, events, and decisions. |
Wikis and docs |
Living documentation in Confluence, SharePoint, or similar platforms. |
File stores |
Files in Google Drive or similar systems. These are indexed where they live and fetched through the source, never copied in bulk. |
Work systems |
Tickets, comments, and status updates from Jira, ServiceNow, or similar tools. Each source keeps its own permissions, honored at fetch time. |
|
A transport protocol is not a source. For example, MCP is how Slack content arrives, not what it is. Every source flows through the same learning pipeline. |
Learning
The learning system determines how the context layer absorbs knowledge from each source. It operates in three modes:
- Index
-
The system saves a pointer to the source and fetches the content at answer time. Nothing is copied.
Example: Point to a returns policy page in Confluence. The page is fetched when a question needs it.
- Extract
-
The system lifts definitions that are already explicit in a source.
Example: Business terms and their meanings, extracted from a glossary document.
- Distill
-
The system infers context from how artifacts are built and used.
Example: A Liveboard’s formulas and filters are distilled into metric definitions and default analysis rules.
The effectiveness of the learning system depends on two factors:
-
Coverage: How much of each source is examined. For a Liveboard, this might include titles, formulas, charts, filters, and drill paths.
-
Efficiency: How well patterns are recognized. Forty similar filters can become one organization-wide rule when the pattern is identified.
Context orchestration
Context orchestration is the process that turns raw learnings into structured, findable context. When a new piece of knowledge arrives, the orchestrator makes four decisions:
| Step | Decision | Example |
|---|---|---|
Allocate |
Which part of the context layer the knowledge belongs in. |
A revenue definition is filed as a metric definition. |
Scope |
The reach of the knowledge: organization, group, data model, or user. |
The definition applies to the Sales data model. |
Reconcile |
Whether to create a new entry, merge with an existing one, or flag a conflict. |
The new ARR definition is merged with the existing one. |
Stamp |
Attach the source, timestamp, and authority level. |
Source: Liveboard |
Structured context is what makes the rest of the system work. Typed, scoped, and stamped context entries are what the RAG layer can rank, the lifecycle system can maintain, and the governance system can audit.