genie-space-patterns

# Genie Space Patterns ## Overview This skill provides patterns for setting up production-ready Databricks Genie Spaces with natural language analytics capabilities. The quality of Genie responses directly correlates with the depth of business context provided in agent instructions. **Core Principle:** Business context drives AI quality. Comprehensive agent instructions, properly selected data assets, and validated benchmark questions ensure reliable Genie performance. --- ## When to Use This Skill Use this skill when: - Creating new Genie Spaces for natural language analytics - Configuring agent behavior and instructions - Selecting and organizing data assets (Metric Views, TVFs, Tables) - Writing benchmark questions for validation - Troubleshooting Genie query routing issues - Optimizing Genie Space performance ### 🔀 Hand Off to `genie-space-export-import-api` Skill When: | User Says / Task Involves | Load Instead | |---|---| | "deploy Genie Space via API" | `genie-space-export-import-api` | | "export Genie Space", "download Genie Space config" | `genie-space-export-import-api` | | "import Genie Space", "restore Genie Space" | `genie-space-export-import-api` | | "CI/CD for Genie Spaces" | `genie-space-export-import-api` | | "migrate Genie Space to another workspace" | `genie-space-export-import-api` | | "back up Genie Space configuration" | `genie-space-export-import-api` | | "programmatically create Genie Space from JSON" | `genie-space-export-import-api` | | "`serialized_space`", "REST API", "`/api/2.0/genie/spaces`" | `genie-space-export-import-api` | **This skill** covers *what* goes into a Genie Space (instructions, assets, benchmarks). The **export/import API skill** covers *how* to deploy it programmatically. --- ## Upstream: Genie API Updates The upstream `databricks-genie` skill provides these MCP tools: | Tool | Purpose | |------|---------| | `list_genie` | List all Genie Spaces accessible to you | | `create_or_update_genie` | Create or update a Genie Space | | `get_genie` | Get Genie Space details | | `delete_genie` | Delete a Genie Space | | `find_genie_by_name` | Look up a Genie Space by name (when you don't have the space_id) | | `ask_genie` | Ask a question to a Genie Space, get SQL + results | | `ask_genie_followup` | Ask follow-up question in existing conversation | **IMPORTANT:** There is NO system table for Genie spaces (e.g., `system.ai.genie_spaces` does NOT exist). To find a Genie space by name, use the `find_genie_by_name` tool. --- ## Critical Rules ### 1. General Instructions Must Be ≤20 Lines **⚠️ CRITICAL:** Genie processes General Instructions effectively only when ≤20 lines. Longer instructions get truncated or ignored. **✅ DO:** Keep General Instructions concise and focused on essential routing rules. **❌ DON'T:** Exceed 20 lines in General Instructions section. ### 2. Benchmark Questions Must Have Working SQL **Every benchmark question MUST include copy-paste-ready SQL that actually runs.** **✅ DO:** Include tested SQL with every benchmark question. **❌ DON'T:** Provide questions without SQL or untested SQL. **⚠️ Temporal Expression Warning:** Avoid `CURRENT_DATE()`, `CURRENT_TIMESTAMP()`, or `DATE_TRUNC('month', CURRENT_DATE)` in benchmark SQL. These produce different results each day, making automated regression testing unreliable. **❌ Fragile (non-deterministic):** ```sql WHERE transaction_date >= DATE_TRUNC('month', CURRENT_DATE) -- Different results every day ``` **✅ Stable for regression testing:** ```sql WHERE transaction_date BETWEEN DATE '2026-01-01' AND DATE '2026-03-31' -- Pinned date range ``` **Guidance:** Use `CURRENT_DATE - 30` for initial interactive testing, but pin to fixed date ranges (`DATE '...'`) for CI/CD regression suites and benchmark validation scripts. ### 3. MEASURE() Uses Column Names, NOT Display Names **The MEASURE() function requires actual column `name`, NOT `display_name`.** **❌ WRONG:** ```sql MEASURE(`Total Revenue`) -- ❌ FAILS: "Total Revenue" is display_name ``` **✅ CORRECT:** ```sql MEASURE(total_revenue) -- ✅ Uses actual column name from YAML ``` **⚠️ Dimension references in metric view queries must use bare names — NOT source/join table prefixes:** **❌ WRONG:** ```sql SELECT dim_store.state_name, MEASURE(total_revenue) FROM ${catalog}.${gold_schema}.revenue_metrics WHERE dim_store.state_name = 'California' -- ❌ UNRESOLVED_COLUMN GROUP BY dim_store.state_name ``` **✅ CORRECT:** ```sql SELECT state_name, MEASURE(total_revenue) FROM ${catalog}.${gold_schema}.revenue_metrics WHERE state_name = 'California' -- ✅ Bare dimension name GROUP BY state_name ``` Metric views flatten all dimensions into a single namespace. When writing benchmark SQL or General Instructions examples, always use bare dimension names without table prefixes. ### 4. Full UC 3-Part Namespace Required **All table and function references MUST use full Unity Catalog namespace.** **❌ WRONG:** ```sql SELECT * FROM fact_sales; SELECT * FROM get_revenue_by_period('2024-01-01', '2024-12-31', 'week'); ``` **✅ CORRECT:** ```sql SELECT * FROM ${catalog}.${gold_schema}.fact_sales; SELECT * FROM ${catalog}.${gold_schema}.get_revenue_by_period('2024-01-01', '2024-12-31', 'week'); ``` ### 5. Data Asset Hierarchy: Metric Views → TVFs → Tables **Always add assets in this order:** 1. **Metric Views** (Primary - use first) - Pre-aggregated, optimized, rich semantics - Best for broad analytical queries 2. **TVFs** (Secondary - use for specific patterns) - Parameterized queries, business logic - Date-bounded queries, top N rankings 3. **Tables** (Last resort - use sparingly) - Only when metric views/TVFs insufficient - Reference data, ad-hoc exploration ### 6. Avoid Contradictory Routing Rules **Issue:** Contradictory rules cause Genie to randomly select wrong assets. **✅ DO:** Group by question type, not asset ```markdown Revenue/booking questions: - By property → revenue_analytics_metrics - By host → get_host_performance TVF (not metric view!) ``` **❌ DON'T:** Create conflicting asset mappings ```markdown - host_analytics_metrics → for host data - get_host_performance → for host data # ❌ CONFLICT! ``` ### 7. Define Ambiguous Terms Explicitly **Common ambiguous terms:** "underperforming", "top performing", "valuable customers", "best hosts" **✅ DO:** Add explicit definitions ```markdown ## Term Definitions "underperforming" = properties with revenue below median (use get_underperforming_properties TVF) "top performing" = highest revenue unless "rated" specified ``` ### 8. TVF Syntax Rules **Common errors to prevent:** **❌ WRONG:** ```sql SELECT * FROM TABLE(get_customer_segments(...)) -- ❌ Don't wrap in TABLE() SELECT * FROM get_customer_segments() -- ❌ Missing parameters SELECT * FROM get_customer_segments(...) GROUP BY segment -- ❌ Unnecessary GROUP BY ``` **✅ CORRECT:** ```sql SELECT * FROM get_customer_segments('2020-01-01', '2024-12-31') -- ✅ Direct call with params ``` ### 9. 🔴 MANDATORY: Serverless SQL Warehouse Only **ALWAYS assign a Serverless SQL Warehouse to Genie Spaces. NEVER use Classic or Pro warehouses.** Serverless provides auto-scaling, instant startup, and cost-efficient idle timedowns -- critical for interactive Genie sessions where users expect sub-10-second responses. **❌ WRONG:** Classic SQL warehouse with manual cluster sizing. **✅ CORRECT:** Serverless SQL warehouse (auto-detected or explicitly set). ### 10. Table/Column COMMENTs Are Genie Fuel **Genie uses Unity Catalog TABLE and COLUMN comments to understand data.** Missing comments = degraded SQL generation quality. **🔴 MANDATORY:** Before adding ANY table as a trusted asset, verify it has: - `COMMENT ON TABLE` with a business-friendly description - `COMMENT ON COLUMN` for every column, including dimension values and business context See [Table Documentation Skill](../../gold/design-workers/06-table-documentation/SKILL.md) for comment standards. **❌ WRONG:** ```sql CREATE TABLE fact_sales (sale_id BIGINT, amt DECIMAL(18,2)); -- No comments, cryptic names ``` **✅ CORRECT:** ```sql CREATE TABLE fact_sales ( sale_id BIGINT COMMENT 'Unique sale identifier from POS system', total_amount DECIMAL(18,2) COMMENT 'Net sale amount in USD after discounts' ) COMMENT 'Daily retail sales transactions at store-SKU grain'; ``` ### 11. Pre-Creation Table Inspection Is Mandatory **Before creating a Genie Space, ALWAYS inspect target table schemas.** Do not rely on assumed schemas. 1. Run `DESCRIBE TABLE EXTENDED` or use `get_table_details` for each trusted asset 2. Verify all tables have TABLE and COLUMN comments 3. Verify descriptive column names (use `customer_lifetime_value` NOT `clv`) 4. Verify proper data types (DATE columns for time-based queries) See [Configuration Guide](references/configuration-guide.md#pre-creation-table-inspection) for the full inspection checklist. ### 12. Prompt User for Benchmark Questions Before Generating **Always ask the user for benchmark questions before generating synthetic ones.** User-provided questions reflect real business needs and catch domain-specific edge cases that synthetic generation misses. **Three outcomes:** - **User provides 10+:** Validate each one. Report any that can't be answered (missing table, ambiguous terms). Proceed with valid set. - **User provides 1-9:** Validate provided, report issues, augment with synthetic to reach 10-15 total. Show augmentation to user. - **User provides none:** Generate 10-15 synthetic benchmarks from asset metadata. Show to user for review. **If a user question can't be answered**, do NOT silently drop it. Inform the user with the specific reason: - "Table `X` is not a trusted asset in this space" - "No data available for churn analysis — available domains are: revenue, bookings, property performance" - "Term 'underperforming' is ambiguous — how should it be defined?" See [Benchmark Intake Workflow](references/benchmark-intake-workflow.md) for the full validation and generation pipeline. ### 13. Validate Programmatically via Conversation API **After deployment, test benchmark questions programmatically using the Conversation API -- not just the UI.** ```python # ✅ Programmatic validation (reproducible, automated) result = ask_genie(space_id="your_space_id", question="What were total sales last month?") assert result["status"] == "COMPLETED" assert result["row_count"] > 0 ``` **Key rules:** - Start a NEW conversation for each unrelated benchmark question - Use `ask_genie_followup` ONLY for related follow-up questions within the same topic - Set timeouts: simple queries (30s), complex joins (60-120s), large scans (120s+) See [Configuration Guide](references/configuration-guide.md#conversation-api-validation) for full testing patterns. ### 14. Extended Instructions Must Follow the 13-Section Structure **⚠️ CRITICAL:** For domains beyond simple use cases, structure Extended Instructions into these 13 mandatory sections: **PURPOSE → ASSET ROUTING → BUSINESS DEFINITIONS → DISAMBIGUATION → AGGREGATION RULES → FUNCTION ROUTING → JOIN GUIDANCE → QUERY RULES → QUERY PATTERNS → TEMPORAL FILTERS → DATA QUALITY NOTES → CONSTRAINTS → SQL EXPRESSIONS** Each section serves a distinct purpose for Genie's SQL generation. Missing sections (especially ASSET ROUTING, DISAMBIGUATION, and BUSINESS DEFINITIONS) are the top causes of misrouted queries. Section 13 (SQL EXPRESSIONS) documents which concepts to promote into structured `sql_snippets`. **✅ DO:** Follow the 13-section template in [Agent Instructions Guide](references/agent-instructions.md#the-13-mandatory-sections). **❌ DON'T:** Write instructions as an unstructured wall of text or a flat numbered list without section headers. ### 15. Column Configs Require Genie-Specific Flags **Every column_config entry should include `enable_format_assistance` and/or `enable_entity_matching` flags based on column type.** | Column Type | enable_format_assistance | enable_entity_matching | |---|---|---| | Dimension text/name | ✅ | ✅ | | Categorical flag (Y/N) | ✅ | ✅ | | Date/timestamp | ✅ | ❌ | | Numeric ID | ✅ | ❌ | | Numeric measure | ❌ | ❌ | See [Agent Instructions Guide](references/agent-instructions.md#column-config-flags-for-genie) for the full pattern. ### 16. Synonyms Go in Column Configs, Not Table COMMENTs **Synonyms belong in Genie Space `column_configs[].synonyms` or metric view YAML `synonyms` fields.** Never embed synonyms in Unity Catalog TABLE or COLUMN COMMENT strings. - UC COMMENTs → business definitions, grain, valid values - column_configs synonyms → user-friendly alternative names for NLQ matching ### 17. Append Instructions When Optimizing — Never Replace **When optimizing an existing Genie Space, APPEND new rules to the existing instruction block. NEVER replace the entire block.** Existing rules were validated against benchmarks. Replacing them risks regression on questions that currently work. See [Agent Instructions Guide](references/agent-instructions.md#instruction-evolution-append-only-updates). ### 18. Validate Programmatically via Conversation API **After deployment, test benchmark questions programmatically using the Conversation API — not just the UI.** ```python result = ask_genie(space_id="your_space_id", question="What were total sales last month?") assert result["status"] == "COMPLETED" assert result["row_count"] > 0 ``` **Key rules:** - Start a NEW conversation for each unrelated benchmark question - Use `ask_genie_followup` ONLY for related follow-up questions within the same topic - Set timeouts: simple queries (30s), complex joins (60-120s), large scans (120s+) See [Configuration Guide](references/configuration-guide.md#conversation-api-validation) for full testing patterns. ### 19. SQL Expressions Provide Structured Business Concept Definitions **SQL Expressions (`sql_snippets` in the API) give Genie structured, parseable definitions of measures, filters, and dimensions.** Unlike text instructions (free-form) and example SQL queries (full query templates), SQL expressions define **individual reusable concepts** that Genie can match directly to user questions. | Type | JSON Key | SQL Requirement | Example | |---|---|---|---| | **Measure** | `measures` | Aggregation function | `SUM(table.total_sales_usd)` | | **Filter** | `filters` | Boolean condition | `table.country_code = 'US'` | | **Dimension** | `expressions` | Column reference or derivation (no aggregation) | `table.zone_combination` | **When to use SQL Expressions:** - KPIs that users ask about frequently (promote from BUSINESS DEFINITIONS) - Common WHERE clauses that appear in 3+ benchmark queries (promote from AGGREGATION RULES) - Grouping attributes with synonyms (promote from DISAMBIGUATION section) **Every SQL Expression MUST include:** - `display_name`: User-friendly name shown in Genie UI - `sql`: Working SQL fragment referencing trusted asset `table.column`