falkordb-snowflake-native-app-skill

name	falkordb-snowflake-native-app-skill
description	Help users install, configure, load data into, query, size, troubleshoot, and use the FalkorDB Snowflake Native App with Cypher, Snowpark Container Services, and Cortex Agents.
allowed-tools	["sql","text"]

FalkorDB Snowflake Native App Skill

Use this skill when the user asks about FalkorDB inside Snowflake, Snowflake Native Apps, graph loading, Cypher, SPCS compute pools, warehouse sizing, FalkorDB resource profiles, or FalkorDB Cortex Agent setup.

Core concepts

• The Snowflake warehouse runs SQL, procedures, staging, and Cortex Agent tool execution.
• The Snowpark Container Services compute pool runs the FalkorDB container.
• FalkorDB container resource options control the request/limit inside the compute pool.
• The app creates an XSMALL warehouse and CPU_X64_S compute pool by default if they do not already exist.
• The default FalkorDB container resources request 1 CPU / 2GB RAM and limit at 2 CPU / 4GB RAM.
• graph_query can write Cypher query results into durable Snowflake tables when called with write.outputTable options.
• The Snowflake Agent can use text_to_cypher for difficult natural-language graph questions before calling run_cypher.
• The Agent should show generated and executed Cypher in user-facing answers.

Standard setup flow

CALL <app_instance_name>.app_public.start_app(
  'FALKORDB_POOL',
  'FALKORDB_WH'
);

CALL <app_instance_name>.app_public.get_service_status();
CALL <app_instance_name>.app_public.graph_list();

To open the FalkorDB Browser, which lets users visually explore graphs, inspect nodes and relationships, and run Cypher queries interactively:

SHOW ENDPOINTS IN SERVICE <app_instance_name>.app_public.st_spcs;

SELECT "ingress_url" AS browser_url
FROM TABLE(RESULT_SCAN(LAST_QUERY_ID()))
WHERE "name" = 'falkordb-browser';

Open the returned browser_url in a web browser after the service is READY.

For custom FalkorDB container allocation:

CALL <app_instance_name>.app_public.stop_app();
CALL <app_instance_name>.app_public.start_app(
  'FALKORDB_POOL',
  'FALKORDB_WH',
  OBJECT_CONSTRUCT(
    'cpuRequest', 2,
    'memoryRequest', '4G',
    'cpuLimit', 4,
    'memoryLimit', '8G'
  )
);

Explain resource options as:

• cpuRequest / memoryRequest: reserved resources Snowflake uses for scheduling the container on the compute pool node.
• cpuLimit / memoryLimit: maximum resources the FalkorDB container can use.
• Requests must fit the selected compute pool; if they are too large, Snowflake can fail scheduling or report insufficient resources.

Loading data

The app loads data from the bound consumer_data_table reference.

CALL <app_instance_name>.app_public.load_csv(
  'my_graph',
  'LOAD CSV FROM ''file://consumer_data.csv'' AS row
   MERGE (n:Node {id: row[0]})
   SET n.name = row[1]'
);

Prefer MERGE for retry-safe loads. For large loads, create an index before loading:

CALL <app_instance_name>.app_public.graph_query(
  'my_graph',
  'CREATE INDEX ON :Node(id)'
);

Querying

CALL <app_instance_name>.app_public.graph_query(
  'my_graph',
  'MATCH (n) RETURN count(n) AS node_count'
);

To persist query results in Snowflake:

CALL <app_instance_name>.app_public.graph_query(
  'my_graph',
  'MATCH (n) RETURN n.name AS name LIMIT 100',
  OBJECT_CONSTRUCT(
    'write', OBJECT_CONSTRUCT(
      'outputTable', 'RESULT_DB.RESULT_SCHEMA.GRAPH_QUERY_RESULTS'
    )
  )
);

Cortex Agent

Create the agent only after the FalkorDB service is running and graphs are loaded.

USE WAREHOUSE FALKORDB_WH;

CALL <app_instance_name>.graph.create_agent(
  'FALKORDB_GRAPH_AGENT',
  'SOURCE_DB.SOURCE_SCHEMA',
  'WORKING_DB.WORKING_SCHEMA'
);

The consumer role must have:

GRANT DATABASE ROLE SNOWFLAKE.CORTEX_AGENT_USER TO ROLE <consumer_role>;
GRANT DATABASE ROLE SNOWFLAKE.CORTEX_USER TO APPLICATION <app_instance_name>;
GRANT IMPORTED PRIVILEGES ON DATABASE SNOWFLAKE TO APPLICATION <app_instance_name>;

Then use Snowflake UI: AI & ML -> Agents or Snowflake Intelligence.

The deployed agent can load data only from the already-bound consumer_data_table reference. It should ask for a graph name if missing, generate LOAD CSV FROM 'file://consumer_data.csv' AS row ... mapping Cypher from the table columns, prefer MERGE, and ask for confirmation before running the load tool.

For difficult graph questions, the deployed agent should call text_to_cypher(input_agent_name, graph_name, user_question) before run_cypher. The tool uses the default Snowflake Cortex model claude-4-sonnet and FalkorDB graph schema context to generate Cypher, and returns the generated cypher, selected model, and the schema_context used. If the user explicitly asks for a specific Cortex model, pass it as the optional fourth argument: text_to_cypher(input_agent_name, graph_name, user_question, model_name). The schema context is labels, relationship types, property keys, and graph stats from the selected FalkorDB graph, not the Snowflake source schema. The role using the agent should have SNOWFLAKE.CORTEX_AGENT_USER; the application should have SNOWFLAKE.CORTEX_USER and imported privileges on the SNOWFLAKE database.

When answering query requests, show the Cypher query. text_to_cypher returns the generated cypher, and run_cypher returns both the executed cypher_query and result.