Writing Context

After building context through scanning and ingestion, you'll want to refine it — edit semantic sources to match your business logic, add knowledge pages that capture tribal knowledge, and query your data through the semantic layer to verify everything works.

Semantic Sources

Semantic sources are YAML files that describe your tables, columns, measures, and joins. They're the core of the context layer — the structured definitions that agents use to generate correct SQL.

Listing sources

# List all sources across connections
ktx sl list

# List sources for a specific connection
ktx sl list --connection-id my-postgres

# Output as JSON
ktx sl list --json

Reading a source

ktx sl read orders --connection-id my-postgres

This prints the full YAML definition for the source.

The source schema

A semantic source defines a single queryable entity — usually a table or a SQL expression. Here's a fully annotated example:

name: orders
description: Customer orders with line-item totals
table: public.orders          # or use `sql:` for a custom SQL expression
grain:
  - order_id                  # columns that uniquely identify a row

columns:
  - name: order_id
    type: string              # string | number | time | boolean
    description: Unique order identifier

  - name: order_date
    type: time
    role: time                # marks this as the default time dimension
    description: Date the order was placed

  - name: status
    type: string
    visibility: public        # public (default) | internal | hidden
    description: Current order status

  - name: _etl_loaded_at
    type: time
    visibility: hidden        # hidden columns are excluded from agent queries
    description: Internal ETL timestamp

  - name: total_amount
    type: number
    description: Order total in USD

measures:
  - name: total_revenue
    expr: SUM(total_amount)
    description: Sum of all order values
  - name: order_count
    expr: COUNT(DISTINCT order_id)
    description: Number of distinct orders
  - name: avg_order_value
    expr: AVG(total_amount)
    description: Average order value
  - name: high_value_revenue
    expr: SUM(total_amount)
    filter: total_amount > 100
    description: Revenue from orders over $100

segments:
  - name: us_orders
    expr: country = 'US'
    description: Orders from US customers

joins:
  - to: customers
    on: orders.customer_id = customers.customer_id
    relationship: many_to_one   # many_to_one | one_to_many | one_to_one
  - to: order_items
    on: orders.order_id = order_items.order_id
    relationship: one_to_many
    alias: items                # optional alias for the joined source

Key fields:

Field	Required	Description
`name`	Yes	Source identifier (lowercase, underscores)
`table` or `sql`	Yes	Database table or custom SQL expression (exactly one)
`grain`	Yes	Columns that define row uniqueness
`columns`	No	Column definitions with type, role, visibility
`measures`	No	Aggregation expressions (SUM, COUNT, AVG, etc.)
`joins`	No	Relationships to other sources
`segments`	No	Named filter conditions
`inherits_columns_from`	No	Inherit column metadata from a manifest entry

Column visibility controls what agents see:

Visibility	Behavior
`public`	Included in agent queries and listings (default)
`internal`	Available for joins and measures but not shown to agents
`hidden`	Excluded entirely — useful for ETL columns

Writing a source

ktx sl write orders --connection-id my-postgres --yaml '
name: orders
table: public.orders
grain: [order_id]
columns:
  - name: order_id
    type: string
  - name: total_amount
    type: number
measures:
  - name: total_revenue
    expr: SUM(total_amount)
'

You can also edit source files directly — they live at semantic-layer/<connection-id>/<source-name>.yaml in your project directory.

Validating sources

Validation checks a source definition against the actual database schema:

ktx sl validate orders --connection-id my-postgres

This catches mismatches — columns that don't exist in the table, type mismatches, invalid join targets — before an agent tries to use the source.

Querying

The semantic layer compiles your measures and dimensions into SQL, optionally executing it against the database:

# Compile a query to SQL
ktx sl query \
  --connection-id my-postgres \
  --measure total_revenue \
  --measure order_count \
  --dimension "order_date" \
  --filter "status = 'completed'" \
  --order-by order_date:desc \
  --limit 10 \
  --format sql

This outputs the compiled SQL without executing it. To run the query:

# Execute and return results
ktx sl query \
  --connection-id my-postgres \
  --measure total_revenue \
  --dimension "order_date" \
  --execute \
  --max-rows 100

Query flags:

Flag	Description
`--measure <name>`	Measure to query (repeatable, at least one required)
`--dimension <name>`	Dimension to group by (repeatable)
`--filter <expr>`	Filter expression (repeatable)
`--segment <name>`	Named segment to apply (repeatable)
`--order-by <field[:dir]>`	Sort field, optionally with `:asc` or `:desc` (repeatable)
`--limit <n>`	Maximum rows in the compiled query
`--format <mode>`	Output format: `json` (default) or `sql`
`--execute`	Execute the query against the database
`--max-rows <n>`	Maximum rows to return when executing
`--include-empty`	Include empty/null rows in results

The query planner is grain-aware — it understands the cardinality of joins and avoids chasm traps (double-counting caused by many-to-many fan-outs). When you query measures that span multiple sources, KTX generates sub-queries at the correct grain before joining.

Knowledge Pages

Knowledge pages are Markdown files that capture business context — definitions, rules, gotchas, and anything an agent needs to understand beyond what the schema tells it.

What they are

When an agent asks "what counts as an active user?" or "why do revenue numbers differ between the dashboard and the SQL query?", the answer isn't in the schema. It's tribal knowledge that lives in Slack threads, Notion pages, or someone's head. Knowledge pages make that context searchable and available to agents.

Organization

Knowledge pages are organized by scope:

knowledge/
├── global/                          # Cross-cutting definitions
│   ├── order-status-definitions.md
│   ├── revenue-recognition-rules.md
│   └── data-freshness-sla.md
└── user/
    └── local/                       # User-scoped context
        ├── schema-conventions.md
        └── known-data-issues.md

Global pages apply across all connections — business definitions, metric standards, company terminology.
User-scoped pages are private to a user ID — personal notes, local gotchas, or context you do not want shared globally.

Writing pages

ktx wiki write order-status-definitions \
  --scope global \
  --summary "Business definitions for order status values" \
  --content "## Order Statuses

- **pending**: Order placed but not yet processed
- **confirmed**: Payment received, awaiting fulfillment
- **shipped**: Order dispatched to carrier
- **delivered**: Order received by customer
- **cancelled**: Order cancelled before shipment

Orders in pending status for more than 48 hours are flagged for review." \
  --tag orders \
  --tag definitions \
  --sl-ref orders

Write flags:

Flag	Description
`--scope <scope>`	`global` (default) or `user`
`--summary <text>`	Short description for search results (required)
`--content <text>`	Full Markdown content (required)
`--tag <tag>`	Categorization tag (repeatable)
`--ref <ref>`	Reference to external resources (repeatable)
`--sl-ref <ref>`	Link to a semantic source (repeatable)

You can also create and edit knowledge pages directly as Markdown files in the knowledge/ directory.

Listing pages

ktx wiki list

Reading a page

ktx wiki read order-status-definitions

Searching

ktx wiki search "revenue recognition"

Search uses both full-text matching and semantic similarity — it finds relevant pages even when the exact terms don't match. Agents call this automatically when they need business context to answer a question.

On this page