app.py

import os
import torch
import requests
import numpy as np
import pandas as pd
import gradio as gr
import google.generativeai as genai
from sentence_transformers import SentenceTransformer
from torch.nn.functional import cosine_similarity

df = pd.read_csv("cleaned_data.csv")

bert_model = SentenceTransformer('all-MiniLM-L6-v2')
df["course_embedding"] = df["Transformed_description"].apply(lambda x: bert_model.encode(x, convert_to_tensor=True))
genai.configure(api_key=os.getenv("GEMINI_API_KEY"))

def recommend_courses(skills, interests, experience, education, time, certificates, careerpath):
    try:
        # Create initial user profile
        user_profile = {
            "skills": [s.strip() for s in skills.split(",") if s.strip()],
            "interests": [s.strip() for s in interests.split(",") if s.strip()],
            "experience": [s.strip() for s in experience.split(",") if s.strip()],
            "education": [s.strip() for s in education.split(",") if s.strip()],
            "time": [s.strip() for s in time.split(",") if s.strip()],
            "certificates": [s.strip() for s in certificates.split(",") if s.strip()],
            "careerpath": [s.strip() for s in careerpath.split(",") if s.strip()]
        }

        # Get skill level assessment
        response = client.models.generate_content(
            model="gemini-pro",
            contents=f"""
            Give the current skill level in one word out of 'beginner', 'intermediate', 'advanced'.
            Here is the user profile: {user_profile}
            strictly do not output any extra textual data."""
        )
        
        CurrentSkill = response.text.strip().replace("\n", "")
        user_profile["CurrentSkill"] = [CurrentSkill]

        user_text = " ".join(user_profile["skills"] + user_profile["interests"] + user_profile["experience"] + user_profile["education"] + user_profile["time"] + user_profile["certificates"] + user_profile["careerpath"] + user_profile["CurrentSkill"])

        # # Create weighted user text representation
        # user_text = " ".join([
        #     " ".join(user_profile["skills"]) * 3,
        #     " ".join(user_profile["interests"]) * 2,
        #     " ".join(user_profile["careerpath"]) * 2,
        #     " ".join(user_profile["experience"]),
        #     " ".join(user_profile["education"]),
        #     " ".join(user_profile["certificates"]),
        #     " ".join(user_profile["CurrentSkill"]) * 2
        # ])

        user_embedding = bert_model.encode(user_text, convert_to_tensor=True)
        course_embeddings = torch.stack(df["course_embedding"].tolist())
        similarities = cosine_similarity(user_embedding, course_embeddings)
        # similarities = cosine_similarity(user_embedding.unsqueeze(0), course_embeddings)[0]

        # Original weighting scheme

        weights = {
            "similarity": 0.6,
            "rating": 0.2,
            "difficulty": 0.1,
            "time_to_complete": 0.1
        }

        df["normalized_rating"] = (df["course_rating"] - df["course_rating"].min()) / (
            df["course_rating"].max() - df["course_rating"].min())
        df["normalized_difficulty"] = 1 - (df["course_difficulty"] / df["course_difficulty"].max())

        df["ranking_score"] = (
            weights["similarity"] * similarities.cpu().numpy() +
            weights["rating"] * df["normalized_rating"].values +
            weights["difficulty"] * df["normalized_difficulty"].values
        )

        top_courses = df.sort_values(by="ranking_score", ascending=False).head(6)
        output = top_courses["course_name"].tolist()

        response2 = client.models.generate_content(
            model="gemini-pro",
            contents=f"""
            Return a JSON object with this exact structure:
            {{
                "beginner": [
                    {{"name": "course name", "url": "course url"}}
                ],
                "intermediate": [
                    {{"name": "course name", "url": "course url"}}
                ],
                "advanced": [
                    {{"name": "course name", "url": "course url"}}
                ]
            }}
            
            Categorize these courses: {output}

            Add Url of the specific course from {df["course_url"]}
            Based on:
            - User skill level: {CurrentSkill}
            - Course difficulties: {top_courses['normalized_difficulty'].tolist()}
            - User skills: {user_profile['skills']}
            
            Categorise atleast one course for each beginner, intermediate and advanced.
            Return ONLY valid JSON without any extra text.
            """
        )

        try:
            json_response = json.loads(response2.text.strip().replace('```json', '').replace('```', ''))
            # Validate structure
            for level in ['beginner', 'intermediate', 'advanced']:
                if level not in json_response:
                    json_response[level] = []
                else:
            # Ensure each course has name and url
                    for course in json_response[level]:
                        if not isinstance(course, dict) or 'name' not in course or 'url' not in course:
                            json_response[level] = []
                        break
            return json_response
        except:
            return {
                "beginner": [],
                "intermediate": [],
                "advanced": [],
                "error": "Failed to categorize courses"
            }

    except Exception as e:
        return {"error": str(e)}

# Create Gradio interface
iface = gr.Interface(
    fn=recommend_courses,
    inputs=[
        gr.Textbox(label="Skills", placeholder="python, machine learning"),
        gr.Textbox(label="Interests", placeholder="AI, data science"),
        gr.Textbox(label="Experience", placeholder="2 years python"),
        gr.Textbox(label="Education", placeholder="bachelor's in CS"),
        gr.Textbox(label="Time Available", placeholder="6 months"),
        gr.Textbox(label="Certificates", placeholder="AWS, GCP"),
        gr.Textbox(label="Career Path", placeholder="ML engineer")
    ],
    outputs=gr.JSON(),
    title="Personalized Course Recommender",
    description="Enter your profile details to get course recommendations organized by difficulty level"
)

if __name__ == "__main__":
    iface.launch(share=True)