RLHF and Admin Approvals

Implement human-in-the-loop safety gates and continuous AI alignment within the Arivu SDK.

Arivu provides two distinct human-in-the-loop mechanisms out of the box: Admin Approvals (to protect database integrity) and RLHF (Reinforcement Learning from Human Feedback, to align AI generation quality).

Admin Approval Gate

Pauses destructive queries (UPDATE, DELETE, DROP) until a verified administrator explicitly approves the execution.

RLHF Signals

Collects user sentiment and routes it back into the memory node for telemetry tracking and future prompt optimization.

---

1. The Admin Approval Gate

When the Arivu engine is parameterized with mode="user", the internal LangGraph state machine automatically restricts data-modifying SQL queries.

Instead of running blindly, the query_verifier node delegates control to the admin_approval node. The pipeline halts execution, writes the SQL payload to pending memory, and returns early.

Architecture Flow

1. Detection

   The LLM formulates a mutation query (e.g., DELETE FROM logs WHERE status = 'expired').

2. Pipeline Intervention

   The pipeline identifies the destructive intent and outputs a PipelineResult with pending_approval = True, keeping the user context perfectly preserved.

3. Resolution Verification

   A permitted user executes a /approve command. The pipeline resumes state, running the db_execution node securely.

Implementing Approvals via SDK

If you’re building a custom client (or extending arivu.integrations.BaseIntegration), you manage the pending state manually using Arivu’s Memory singleton:

my_integration.py

from arivu.memory.store import get_pending_approval, resolve_approval

def handle_admin_interaction(user_action: str, session_id: str):
    # 1. Fetching the quarantined SQL
    pending = get_pending_approval(session_id)
    if pending:
        sql_query = pending['sql']
        print("Safety Gate: Requires Approval for ->", sql_query)
    
    # 2. Resolving based on user action
    is_approved = (user_action == "/approve")
    
    # Passes boolean to memory. True routes to DB Execution. False cancels it.
    resolve_approval(session_id, approve=is_approved)

If Arivu.connect() is run with mode="admin", the LangGraph conditionally bypasses the admin_approval node and enforces mutations unconditionally. Use this solely for restricted dev environments.

---

2. Dynamic RLHF Integration

RLHF ensures your Text-to-SQL logic converges strictly onto your users’ intents over time. If a user corrects the bot (e.g., “No, I meant total gross revenue, not net”), capturing that correction prevents identical hallucinations later.

Feeding execution signals

You inject sentiment back into the graph when issuing run_pipeline(). The rlhf_signal can be a binary sentiment or a comprehensive text explanation.

query_handler.py

# 1. Standard execution
app.run_pipeline({
    "question": "Show me our active users", 
    "session_id": "sid_1"
})

# 2. RLHF triggered on correction
result = app.run_pipeline({
    "question": "Group them by region actually.", 
    "session_id": "sid_1"
}, rlhf_signal="negative - bot failed to group by region initially")

print("Pipeline Adjusted:", result.success)

The rlhf_feedback LangGraph node executes entirely downstream of the response_generator. Injecting RLHF signals adds zero blocking latency to your conversational UX.