app.py

import os
import random
import time
import pandas as pd
import numpy as np
import gradio as gr
import torch
from sentence_transformers import SentenceTransformer
import faiss
import requests
from dotenv import load_dotenv

# ========================
# Initialization
# ========================
load_dotenv()

HF_API_TOKEN = (os.getenv("HF_API_TOKEN") or "").strip()
HF_INFERENCE_ENDPOINT = "https://api-inference.huggingface.co/models/distilbert-base-uncased"

# fallback in case the token isn't available
if not HF_API_TOKEN:
    print("⚠️ Warning: No HF_API_TOKEN found — using read-only mode (no inference calls).")

# Vector memory setup
embedder = SentenceTransformer("all-MiniLM-L6-v2")
embedding_dim = 384
index = faiss.IndexFlatL2(embedding_dim)
incident_memory = []  # stores {vector, metadata}

# Helper: create embeddings
def embed_text(text):
    vector = embedder.encode([text], convert_to_numpy=True)
    return vector

# ========================
# Core Functions
# ========================
def detect_anomaly(event):
    """Simple adaptive anomaly detection."""
    # Random anomaly forcing for test verification
    force_anomaly = random.random() < 0.25  # 25% of events become anomalies automatically
    if force_anomaly or event["latency"] > 150 or event["error_rate"] > 0.05:
        return "Anomaly"
    return "Normal"

def analyze_with_hf_api(text):
    """Call Hugging Face Inference API safely."""
    if not HF_API_TOKEN:
        return "⚠️ No API token — running offline simulation."
    headers = {"Authorization": f"Bearer {HF_API_TOKEN}"}
    try:
        response = requests.post(
            HF_INFERENCE_ENDPOINT,
            headers=headers,
            json={"inputs": text},
            timeout=5
        )
        if response.status_code == 200:
            result = response.json()
            if isinstance(result, list):
                return result[0].get("label", "No label")
            return str(result)
        else:
            return f"Error {response.status_code}: {response.text}"
    except Exception as e:
        return f"Error generating analysis: {e}"

def simulate_healing(event):
    """Simulate automated remediation based on anomaly context."""
    actions = [
        "Restarted container",
        "Scaled up pods",
        "Cleared queue backlog",
        "Purged cache and retried"
    ]
    if event["status"] == "Anomaly":
        return random.choice(actions)
    return "-"

def add_to_vector_memory(event):
    """Store event context in FAISS for post-incident learning."""
    text = f"Component: {event['component']} | Latency: {event['latency']} | ErrorRate: {event['error_rate']} | Analysis: {event['analysis']}"
    vector = embed_text(text)
    index.add(vector)
    incident_memory.append({
        "vector": vector,
        "metadata": text
    })
    return len(incident_memory)

def find_similar_events(event, top_k=3):
    """Find semantically similar past incidents."""
    if len(incident_memory) < 3:
        return "Not enough incidents stored yet."
    text = f"Component: {event['component']} | Latency: {event['latency']} | ErrorRate: {event['error_rate']} | Analysis: {event['analysis']}"
    query_vec = embed_text(text)
    distances, indices = index.search(query_vec, top_k)
    results = [incident_memory[i]["metadata"] for i in indices[0] if i < len(incident_memory)]
    return f"Found {len(results)} similar incidents (e.g., {results[0][:100]}...)." if results else "No matches found."

# ========================
# Event Handling
# ========================
events = []

def process_event(component, latency, error_rate):
    event = {
        "timestamp": time.strftime("%Y-%m-%d %H:%M:%S"),
        "component": component,
        "latency": float(latency),
        "error_rate": float(error_rate),
    }

    event["status"] = detect_anomaly(event)
    event["analysis"] = analyze_with_hf_api(f"{component} latency={latency}, error={error_rate}")
    event["healing_action"] = simulate_healing(event)

    # Vector memory & similarity learning
    add_to_vector_memory(event)
    event["healing_action"] += " " + find_similar_events(event)

    events.append(event)
    if len(events) > 20:
        events.pop(0)

    df = pd.DataFrame(events)
    return "✅ Event Processed", df

# ========================
# Gradio UI
# ========================
with gr.Blocks(title="Agentic Reliability Framework MVP") as demo:
    gr.Markdown("## 🧠 Agentic Reliability Framework MVP\nAdaptive anomaly detection + AI-driven self-healing + vector memory")

    with gr.Row():
        component_input = gr.Dropdown(
            ["api-service", "data-ingestor", "queue-worker", "model-runner"],
            label="Component",
            value="api-service"
        )
        latency_input = gr.Number(label="Latency (ms)", value=random.uniform(50, 200))
        error_input = gr.Number(label="Error Rate", value=random.uniform(0.01, 0.15))

    submit_btn = gr.Button("🚀 Submit Telemetry Event")

    output_text = gr.Textbox(label="Detection Output")
    output_table = gr.Dataframe(headers=["timestamp", "component", "latency", "error_rate", "analysis", "status", "healing_action"], label="Recent Events (Last 20)")

    submit_btn.click(
        fn=process_event,
        inputs=[component_input, latency_input, error_input],
        outputs=[output_text, output_table]
    )

# ========================
# Launch
# ========================
if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)