diff --git a/apps/accounts/models.py b/apps/accounts/models.py
index 28428c0..f84df27 100644
--- a/apps/accounts/models.py
+++ b/apps/accounts/models.py
@@ -112,49 +112,28 @@ class Role(IdentifierMixin, TimeStampMixin, Model):
def create_default_agents_for_role(sender, instance: Role, created: bool, **kwargs):
if created:
from apps.onboarding.models import AgentConfig # L: circular import :(
-
+ from apps.onboarding.consumers.prompts import OnboardingPrompts
default_agents = [
{
'type': 'curriculum',
'name': f"{instance.name} Curriculum Agent",
- 'prompt': (
- f"You are an instructional design assistant for onboarding the role '{instance.name}'. "
- "Your job is to teach the learner what the role does and how responsibilities are performed in practice. "
- "Create a structured curriculum with clear objectives, prerequisite knowledge, core competencies, "
- "hands-on tasks, and measurable outcomes. Avoid role-play and avoid claiming to be in the role; "
- "focus on teaching the role responsibilities, expected decisions, and quality standards."
- )
+ 'prompt': OnboardingPrompts.default_curriculum_prompt(instance.name),
},
{
'type': 'knowledge',
'name': f"{instance.name} Knowledge Agent",
- 'prompt': (
- f"You are a domain knowledge tutor for the role '{instance.name}'. "
- "Answer questions with concise explanations, practical examples, and references to expected workflows. "
- "When possible, explain why a step matters, common mistakes, and how to verify correctness. "
- "Do not act as the role holder; teach the learner how to perform the role responsibly and accurately."
- )
+ 'prompt': OnboardingPrompts.default_knowledge_prompt(instance.name),
},
{
'type': 'assessment',
'name': f"{instance.name} Assessment Agent",
- 'prompt': (
- f"You are an assessment designer for onboarding the role '{instance.name}'. "
- "Generate scenario-based checks that evaluate conceptual understanding, decision-making, and execution quality. "
- "Include rubrics, expected evidence, and feedback that explains gaps and remediation steps. "
- "Assess against role responsibilities and standards, not generic trivia."
- )
+ 'prompt': OnboardingPrompts.default_assessment_prompt(instance.name),
},
{
'type': 'monitor',
'name': f"{instance.name} Progress Monitor",
- 'prompt': (
- f"You are a progress coaching assistant for learners training for the role '{instance.name}'. "
- "Track competency milestones, summarize strengths and weaknesses, and recommend next actions. "
- "Flag unresolved risks, missing evidence, and topics requiring revision. "
- "Keep feedback specific, actionable, and tied to role responsibilities and expected outcomes."
- )
- }
+ 'prompt': OnboardingPrompts.default_monitor_prompt(instance.name),
+ },
]
with transaction.atomic():
diff --git a/apps/knowledge/tasks.py b/apps/knowledge/tasks.py
index 7865d01..e4faa02 100644
--- a/apps/knowledge/tasks.py
+++ b/apps/knowledge/tasks.py
@@ -1,4 +1,5 @@
import hashlib
+import logging
from celery import shared_task
from django.conf import settings
@@ -9,6 +10,8 @@ from pypdf import PdfReader
from apps.knowledge.models import RoleRagDocument, TrainingFile
+logger = logging.getLogger(__name__)
+
def _decode_text_bytes(raw_bytes: bytes) -> str:
try:
@@ -41,6 +44,10 @@ def _get_text_chunks(text: str, size: int = 10000):
@shared_task(name="apps.knowledge.tasks.ingest_training_file_task", bind=True, soft_time_limit=900, time_limit=1200)
def ingest_training_file_task(self, file_uuid):
+ """
+ Ingests a training file by extracting text, chunking it, generating embeddings via an external service,
+ and saving RoleRagDocument entries. Updates the file status accordingly and triggers prompt refinement.
+ """
try:
file_obj = TrainingFile.objects.get(uuid=file_uuid)
except TrainingFile.DoesNotExist:
@@ -99,6 +106,10 @@ def ingest_training_file_task(self, file_uuid):
file_obj.status = 'embedded'
file_obj.is_processed = True
file_obj.save()
+
+ if file_obj.role_id:
+ update_agent_prompts_from_file_task.delay(str(file_obj.role.uuid))
+
return f"Processed {chunk_counter} chunks via batching."
except Exception as e:
@@ -106,3 +117,69 @@ def ingest_training_file_task(self, file_uuid):
file_obj.description = str(e)
file_obj.save()
raise e
+
+
+@shared_task(name="apps.knowledge.tasks.update_agent_prompts_from_file_task", bind=True, soft_time_limit=120, time_limit=180)
+def update_agent_prompts_from_file_task(self, role_uuid: str):
+ """
+ After a training file is ingested (or deleted), refine the curriculum AgentConfig
+ system prompt using document content. Resets to the canonical base prompt when no
+ files remain.
+ """
+ from apps.accounts.models import Role
+ from apps.onboarding.consumers.prompts import OnboardingPrompts
+ from apps.onboarding.models import AgentConfig
+
+ try:
+ role = Role.objects.get(uuid=role_uuid)
+ except Role.DoesNotExist:
+ logger.warning("update_agent_prompts_from_file_task: role %s not found", role_uuid)
+ return
+
+ curriculum_config = AgentConfig.objects.filter(role=role, agent_type='curriculum').first()
+ if not curriculum_config:
+ logger.warning("update_agent_prompts_from_file_task: no curriculum config for role %s", role_uuid)
+ return
+
+ chunk_texts = list(
+ RoleRagDocument.objects.filter(role=role, is_active=True)
+ .order_by('training_file_id', 'chunk_index')
+ .values_list('content', flat=True)[:30]
+ )
+
+ # No files left... so we should reset
+ if not chunk_texts:
+ curriculum_config.system_prompt = OnboardingPrompts.default_curriculum_prompt(role.name)
+ curriculum_config.save(update_fields=['system_prompt', 'updated_at'])
+ logger.info("update_agent_prompts_from_file_task: reset to base prompt for role %s", role_uuid)
+ return
+
+ combined_text = '\n\n'.join(chunk_texts)[:6000]
+ base_prompt = OnboardingPrompts.default_curriculum_prompt(role.name)
+
+ try:
+ with Client(timeout=Timeout(60.0)) as client:
+ response = client.post(
+ settings.INFERENCE_CHAT_COMPLETIONS_ENDPOINT,
+ json={
+ "model": "meta-llama-3.1-8b-instruct",
+ "messages": [
+ {
+ "role": "user",
+ "content": OnboardingPrompts.refine_curriculum_prompt(
+ role.name, base_prompt, combined_text
+ ),
+ },
+ ],
+ "max_tokens": 600,
+ },
+ )
+ response.raise_for_status()
+ refined_prompt = response.json()["choices"][0]["message"]["content"].strip()
+ except Exception as e:
+ logger.exception("update_agent_prompts_from_file_task: LLM call failed for role %s: %s", role_uuid, e)
+ return
+
+ curriculum_config.system_prompt = refined_prompt
+ curriculum_config.save(update_fields=['system_prompt', 'updated_at'])
+ logger.info("update_agent_prompts_from_file_task: refined curriculum prompt for role %s", role_uuid)
diff --git a/apps/knowledge/viewsets.py b/apps/knowledge/viewsets.py
index f4219f9..945ad11 100644
--- a/apps/knowledge/viewsets.py
+++ b/apps/knowledge/viewsets.py
@@ -8,6 +8,7 @@ from apps.accounts.models import Organization, Role
from apps.accounts.permissions import can_manage_organization
from apps.knowledge.models import RoleRagDocument, TrainingFile
from apps.knowledge.serializers import RoleRagDocumentSerializer, TrainingFileSerializer
+from apps.knowledge.tasks import update_agent_prompts_from_file_task
class TrainingFileViewSet(ModelViewSet):
queryset = TrainingFile.objects.all()
@@ -89,7 +90,11 @@ class TrainingFileViewSet(ModelViewSet):
if not (is_uploader or is_org_owner or is_org_manager):
raise PermissionDenied('Permission denied')
- return super().destroy(request, *args, **kwargs)
+ role_uuid = str(instance.role.uuid) if instance.role_id else None
+ response = super().destroy(request, *args, **kwargs)
+ if role_uuid:
+ update_agent_prompts_from_file_task.delay(role_uuid)
+ return response
class RoleRagDocumentViewSet(ReadOnlyModelViewSet):
queryset = RoleRagDocument.objects.all()
diff --git a/apps/onboarding/consumers/prompts.py b/apps/onboarding/consumers/prompts.py
index 3c2ad2b..ce84fc5 100644
--- a/apps/onboarding/consumers/prompts.py
+++ b/apps/onboarding/consumers/prompts.py
@@ -70,6 +70,57 @@ class OnboardingPrompts:
f"Progress context JSON:\n{json.dumps(progress_context)}"
)
+ ### Default agent system prompts (canonical source of truth) ###
+
+ @staticmethod
+ def default_curriculum_prompt(role_name: str) -> str:
+ return (
+ f"You are an instructional design assistant for onboarding the role '{role_name}'. "
+ "Your job is to teach the learner what the role does and how responsibilities are performed in practice. "
+ "Create a structured curriculum with clear objectives, prerequisite knowledge, core competencies, "
+ "hands-on tasks, and measurable outcomes. Avoid role-play and avoid claiming to be in the role; "
+ "focus on teaching the role responsibilities, expected decisions, and quality standards."
+ )
+
+ @staticmethod
+ def default_knowledge_prompt(role_name: str) -> str:
+ return (
+ f"You are a domain knowledge tutor for the role '{role_name}'. "
+ "Answer questions with concise explanations, practical examples, and references to expected workflows. "
+ "When possible, explain why a step matters, common mistakes, and how to verify correctness. "
+ "Do not act as the role holder; teach the learner how to perform the role responsibly and accurately."
+ )
+
+ @staticmethod
+ def default_assessment_prompt(role_name: str) -> str:
+ return (
+ f"You are an assessment designer for onboarding the role '{role_name}'. "
+ "Generate scenario-based checks that evaluate conceptual understanding, decision-making, and execution quality. "
+ "Include rubrics, expected evidence, and feedback that explains gaps and remediation steps. "
+ "Assess against role responsibilities and standards, not generic trivia."
+ )
+
+ @staticmethod
+ def default_monitor_prompt(role_name: str) -> str:
+ return (
+ f"You are a progress coaching assistant for learners training for the role '{role_name}'. "
+ "Track competency milestones, summarize strengths and weaknesses, and recommend next actions. "
+ "Flag unresolved risks, missing evidence, and topics requiring revision. "
+ "Keep feedback specific, actionable, and tied to role responsibilities and expected outcomes."
+ )
+
+ @staticmethod
+ def refine_curriculum_prompt(role_name: str, base_prompt: str, document_text: str) -> str:
+ return (
+ f"You are refining a curriculum agent's system prompt for the '{role_name}' role. "
+ "Training documents have been uploaded. Rewrite the system prompt below so it incorporates "
+ "the specific topics and subject matter from those documents. "
+ "Preserve all original instructions and add concrete topic guidance where relevant. "
+ "Return ONLY the refined system prompt text — no commentary, no labels.\n\n"
+ f"Original system prompt:\n{base_prompt}\n\n"
+ f"Training document content:\n{document_text}"
+ )
+
FALLBACK_SYSTEM_PROMPT = 'You are a helpful onboarding assistant.'
KA_HELP_FALLBACK = (
diff --git a/site/src/views/OnboardingView.vue b/site/src/views/OnboardingView.vue
index d6a9226..a23ddfd 100644
--- a/site/src/views/OnboardingView.vue
+++ b/site/src/views/OnboardingView.vue
@@ -113,7 +113,11 @@ const currentPageBody = computed(() => {
const renderedBody = computed(() => {
if (!currentPageBody.value) return ''
- const body = currentPageBody.value.replace(/^#{1,6}\s+.+\n?/, '')
+ const lines = currentPageBody.value.split('\n')
+ const firstLineText = lines[0].replace(/^#{1,6}\s*/, '').trim()
+ const pageTitle = (currentPage.value?.title ?? '').trim()
+ const startsWithTitle = pageTitle && firstLineText.toLowerCase() === pageTitle.toLowerCase()
+ const body = startsWithTitle ? lines.slice(1).join('\n').trimStart() : currentPageBody.value
return DOMPurify.sanitize(marked.parse(body) as string)
})
diff --git a/site/src/views/OrganizationManage.vue b/site/src/views/OrganizationManage.vue
index e46f61b..f7fe8d9 100644
--- a/site/src/views/OrganizationManage.vue
+++ b/site/src/views/OrganizationManage.vue
@@ -969,6 +969,7 @@ onMounted(async () => {
"
:multiple="false"
:auto-upload="false"
+ :file-list="[]"
>
@@ -1047,6 +1048,7 @@ onMounted(async () => {
"
:multiple="false"
:auto-upload="false"
+ :file-list="[]"
>