feat(AI Assistant): generates formulas for workflow nodes (baserow#4199)

* Kuma generates formulas for workflow nodes

* address feedback

* Set mode=BASEROW_FORMULA_MODE_ADVANCED, not type.

---------

Co-authored-by: peter_baserow <peter@baserow.io>

Author: silvestrid

backend/Makefile

@@ -152,7 +152,7 @@ lint: .check-dev
 	$(VISORT) --check --skip generated $(BACKEND_SOURCE_DIRS) $(BACKEND_TESTS_DIRS)
 	# TODO: make baserow command reading dotenv files
 	DJANGO_SETTINGS_MODULE=$(DJANGO_SETTINGS_MODULE) $(VBASEROW) makemigrations --dry-run --check
-	$(VBANDIT) -r --exclude src/baserow/test_utils $(BACKEND_SOURCE_DIRS)
+	$(VBANDIT) -r --exclude src/baserow/test_utils,src/baserow/config/settings/local.py $(BACKEND_SOURCE_DIRS)
 
 lint-python: lint
 

backend/src/baserow/config/settings/base.py

@@ -1049,7 +1049,7 @@ def __setitem__(self, key, value):
 # The minimum amount of minutes the periodic task's "minute" interval
 # supports. Self-hosters can run every minute, if they choose to.
 INTEGRATIONS_PERIODIC_MINUTE_MIN = int(
-    os.getenv("BASEROW_INTEGRATIONS_PERIODIC_MINUTE_MIN", 1)
+    os.getenv("BASEROW_INTEGRATIONS_PERIODIC_MINUTE_MIN") or 1
 )
 
 TOTP_ISSUER_NAME = os.getenv("BASEROW_TOTP_ISSUER_NAME", "Baserow")

enterprise/backend/src/baserow_enterprise/assistant/prompts.py

@@ -1,3 +1,5 @@
+from django.conf import settings
+
 CORE_CONCEPTS = """
 ### BASEROW STRUCTURE
 
@@ -50,21 +52,22 @@
 """
 
 ASSISTANT_SYSTEM_PROMPT = (
-    """
+    f"""
 You are Kuma, an AI expert for Baserow (open-source no-code platform).
 
 ## YOUR KNOWLEDGE
 1. **Core concepts** (below)
 2. **Detailed docs** - use search_docs tool to search when needed
-3. **API specs** - guide users to https://api.baserow.io/api/schema.json
+3. **API specs** - guide users to "{settings.PUBLIC_BACKEND_URL}/api/schema.json"
+4. **Official website** - "https://baserow.io"
 
 ## HOW TO HELP
 • Use American English spelling and grammar
 • Be clear, concise, and actionable
 • For troubleshooting: ask for error messages or describe expected vs actual results
 • **NEVER** fabricate answers or URLs. Acknowledge when you can't be sure.
 • When you have the tools to help, **ALWAYS** use them instead of answering with instructions.
-* At the end, **always** ask follow-up questions to understand user needs and continue the conversation.
+• When finished, briefly suggest one or more logical next steps only if they use tools you have access to and directly builds on what was just done.
 
 ## FORMATTING (CRITICAL)
 • **No HTML**: Only Markdown (bold, italics, lists, code, tables)

enterprise/backend/src/baserow_enterprise/assistant/tools/automation/prompts.py

@@ -0,0 +1,86 @@
+GENERATE_FORMULA_PROMPT = """
+You are a formula builder. Generate formulas using these functions:
+
+**Comparison operators** (for router conditions only):
+equal, not_equal, greater_than, less_than, greater_than_equal, less_than_equal
+- Arguments: numbers, 'strings', or get() functions
+- Returns: boolean
+- Example: greater_than(get('age'), 18)
+
+**concat(...args)** - Joins arguments into a string
+- Arguments: 'string literals' or get() functions
+- Example: concat('Hello ', get('name'), '!')
+
+**get(path)** - Retrieves values from context using path notation
+- Objects: get('user.name')
+- Arrays: get('items.0'), get('orders.2.total')
+- Nested: get('users.0.address.city')
+- All: get('users.*.email') returns a list of emails from all users
+
+**if(condition, true_value, false_value)** - Conditional expression
+- Arguments: a boolean condition, value if true, value if false
+- Example: if(greater_than(get('score'), 50), 'pass', 'fail')
+
+**today()** - Returns the current date
+**now()** - Returns the current date and time
+
+**constants**:
+- A string literal enclosed in single quotes (e.g., 'hello world', '123')
+
+**Example 1 - String Fields:**
+Input:
+fields_to_resolve: {
+    "ai_prompt": "Determine the priority level based on {{ trigger.title }} and {{ trigger.due_date }}. Choices are: High, Medium, Low.",
+}
+context: {"previous_node": {"1": [{"title": "Finish report", "due_date": "2025-11-08"}]}}
+context_metadata: {
+    "1": {"id": 1, "ref": "trigger", "field_1": {"name": "title", "type": "string"}, "field_2": {"name": "due_date", "type": "date"}},
+    "today": "2025-11-07"
+}
+feedback: ""
+Output:
+generated_formula: {
+    "ai_prompt": "concat(
+        'Determine the priority level based on ',
+        get('previous_node.1.0.title'),
+        ' and ',
+        get('previous_node.1.0.due_date'),
+        '. Choices are: High, Medium, Low.'
+    )"
+}
+**Example 2 - Router Conditions:**
+Input:
+fields_to_resolve: {
+    "condition_1": "Check if {{ trigger.amount }} is greater than 1000",
+}
+context: {"previous_node": {"1": [{"amount": 1500}]}},
+context_metadata: {
+    "1": {"id": 1, "ref": "trigger", "field_1": {"name": "amount", "type": "number"}},
+}
+feedback: ""
+Output:
+generated_formula: {
+    "condition_1": "greater_than(get('previous_node.1.0.amount'), 1000)"
+}
+
+**Task:**
+
+You are given:
+* **fields_to_resolve** — a dictionary where each key is a field name and each value contains instructions to generate a formula.
+* **context** — a dictionary containing the available data.
+* **context_metadata** — a dictionary describing the structure and types within the context.
+* **feedback** — optional information with reported formula errors from previous runs.
+
+**Goal:**
+Generate a dictionary called **generated_formula**, where:
+
+* Keys are the field names from **fields_to_resolve**.
+* Values are valid formulas that can be used in the automation node.
+
+**Rules:**
+
+1. Feel free to skip fields whose description starts with `[optional]`.
+2. Exclude any field if you cannot generate a valid formula for it.
+3. If **feedback** is provided, use it to refine or correct the generated formulas.
+4. Strive to produce the most accurate and useful formulas possible based on the provided context and metadata.
+"""

enterprise/backend/src/baserow_enterprise/assistant/tools/automation/tools.py

@@ -58,7 +58,8 @@ def create_workflows(
         automation_id: int, workflows: list[WorkflowCreate]
     ) -> dict[str, Any]:
         """
-        Create one or more workflows in an automation.
+        Create one or more workflows in an automation. Always use {{ node.ref }} to
+        reference previous nodes values inside the workflow.
 
         :param automation_id: The automation application ID
         :param workflows: List of workflows to create
@@ -67,21 +68,26 @@ def create_workflows(
 
         nonlocal user, workspace, tool_helpers
 
-        tool_helpers.update_status(_("Creating workflows..."))
-
         created = []
 
         automation = utils.get_automation(automation_id, user, workspace)
         for wf in workflows:
             with transaction.atomic():
-                orm_workflow = utils.create_workflow(user, automation, wf)
+                orm_workflow, node_mapping = utils.create_workflow(
+                    user, automation, wf, tool_helpers
+                )
                 created.append(
                     {
                         "id": orm_workflow.id,
                         "name": orm_workflow.name,
                         "state": orm_workflow.state,
                     }
                 )
+
+            # In separate transactions, try to update the formulas inside the workflow,
+            # so we don't block the main creation if something goes wrong here.
+            utils.update_workflow_formulas(wf, node_mapping, tool_helpers)
+
         # Navigate to the last created workflow
         tool_helpers.navigate_to(
             WorkflowNavigationType(

enterprise/backend/src/baserow_enterprise/assistant/tools/automation/types/__init__.py

@@ -2,6 +2,7 @@
     AiAgentNodeCreate,
     CreateRowActionCreate,
     DeleteRowActionCreate,
+    HasFormulasToCreateMixin,
     NodeBase,
     RouterNodeCreate,
     SendEmailActionCreate,
@@ -21,4 +22,5 @@
     "SendEmailActionCreate",
     "AiAgentNodeCreate",
     "TriggerNodeCreate",
+    "HasFormulasToCreateMixin",
 ]