app.py

import gradio as gr
import os
from inference import get_evo_response, get_gpt_response
from logger import log_feedback
import csv
import subprocess

# Helper to load Hall of Fame
def load_hall_of_fame():
    entries = []
    if os.path.exists("feedback_log.csv"):
        with open("feedback_log.csv", newline='', encoding='utf-8') as f:
            reader = csv.DictReader(f)
            for row in reader:
                try:
                    score = float(row.get("evo_was_correct", "0") == "yes")
                    if "👍" in row.get("feedback", "") or score > 0.85:
                        entries.append(row)
                except:
                    continue
    return entries[-10:][::-1]  # last 10, reverse order

def handle_query(question, option1, option2, context):
    options = [option1, option2]
    evo_answer, evo_reasoning, evo_score, evo_context = get_evo_response(question, options, context)
    gpt_answer = get_gpt_response(question, context)
    return (
        f"Answer: {evo_answer} (Confidence: {evo_score:.2f})\n\nReasoning: {evo_reasoning}\n\nContext used: {evo_context[:400]}...",
        gpt_answer,
        f"{question} | {context} | {evo_answer}"
    )

def handle_feedback(feedback_text, question, option1, option2, context, evo_output):
    evo_was_correct = "👍" in feedback_text
    log_feedback(question, option1, option2, context, evo_output, evo_was_correct)
    return "✅ Feedback logged and Evo will improve."

def trigger_retrain():
    try:
        subprocess.run(["python", "retrain_from_feedback.py"], check=True)
        return "🔁 Evo retraining completed."
    except subprocess.CalledProcessError:
        return "❌ Retraining failed. Check logs."

def render_hof():
    entries = load_hall_of_fame()
    if not entries:
        return "No Hall of Fame entries yet. Submit feedback!"
    result = "\n\n".join(
        [
            f"🏆 **Q:** {e['question']}\n**A:** {e['evo_output']}\n**Feedback:** {e.get('feedback', 'N/A')}\n**Context:** {e['context'][:200]}..."
            for e in entries
        ]
    )
    return result

description = """
# 🧠 EvoRAG – Adaptive Reasoning AI

**What is Evo?**  
EvoTransformer is a lightweight, evolving neural network with ~28M parameters.  
It learns from feedback, adapts over time, and reasons using both web and context data.  

**Why Evo?**  
✅ Evolves from human input  
✅ Architecturally updatable  
✅ Transparent and fine-tunable  
✅ Efficient on modest hardware

**Hardware**: Trained on Google Colab CPU/GPU  
**Token limit**: 128  
**Benchmark**: PIQA, HellaSwag, ARC  
**Version**: Evo v2.2 (Memory + Web Retrieval + Feedback Learning)
"""

with gr.Blocks(title="EvoRAG") as demo:
    gr.Markdown(description)
    with gr.Row():
        question = gr.Textbox(label="📝 Ask anything", placeholder="e.g., What’s the best way to escape a house fire?")
    with gr.Row():
        option1 = gr.Textbox(label="Option A", placeholder="e.g., Run outside")
        option2 = gr.Textbox(label="Option B", placeholder="e.g., Hide under bed")
    context = gr.Textbox(label="📂 Optional Context", placeholder="Paste any extra background info here", lines=3)

    submit_btn = gr.Button("🔍 Run Comparison")
    with gr.Row():
        evo_output = gr.Textbox(label="🧠 EvoRAG's Reasoned Answer", lines=6)
        gpt_output = gr.Textbox(label="🤖 GPT-3.5's Suggestion", lines=6)

    feedback = gr.Radio(["👍 Evo was correct. Retrain from this.", "👎 Evo was wrong. Don't retrain."], label="Was Evo’s answer useful?", value=None)
    submit_feedback = gr.Button("📬 Submit Feedback")
    feedback_status = gr.Textbox(label="Feedback Status", interactive=False)

    retrain_button = gr.Button("🔄 Retrain Evo Now")
    retrain_status = gr.Textbox(label="Retraining Status", interactive=False)

    with gr.Accordion("🏆 Evo Hall of Fame (Top Reasoning Entries)", open=False):
        hof_display = gr.Markdown(render_hof())

    submit_btn.click(fn=handle_query, inputs=[question, option1, option2, context], outputs=[evo_output, gpt_output, feedback_status])
    submit_feedback.click(
        fn=lambda fb, q, o1, o2, ctx, eo: handle_feedback(fb, q, o1, o2, ctx, eo),
        inputs=[feedback, question, option1, option2, context, feedback_status],
        outputs=[feedback_status]
    )
    retrain_button.click(fn=trigger_retrain, inputs=[], outputs=[retrain_status])

demo.launch(server_name="0.0.0.0", server_port=7860, share=True)