import streamlit as st import pandas as pd import numpy as np # --- CONFIGURATION --- st.set_page_config(page_title="AI Concept Lab", page_icon="๐Ÿงช", layout="wide") # --- STYLING --- st.markdown(""" """, unsafe_allow_html=True) # --- STATE MANAGEMENT --- if 'xp' not in st.session_state: st.session_state.xp = 0 if 'completed_modules' not in st.session_state: st.session_state.completed_modules = set() # --- SIDEBAR CURRICULUM --- with st.sidebar: st.title("๐Ÿงช AI Concept Lab") st.write(f"**XP Gained:** {st.session_state.xp}") st.progress(len(st.session_state.completed_modules) / 3) st.write("---") st.header("๐Ÿ“š Curriculum") page = st.radio("Select Module:", [ "1. The Stochastic Parrot (Temperature)", "2. Meaning in Numbers (Embeddings)", "3. The Knowledge Bridge (RAG)" ]) st.write("---") if st.button("Reset Progress"): st.session_state.xp = 0 st.session_state.completed_modules = set() st.rerun() # --- MODULE 1: TEMPERATURE --- if page == "1. The Stochastic Parrot (Temperature)": st.markdown('
Module 1: Controlling Creativity
', unsafe_allow_html=True) st.caption("Reference: *Prompt Engineering for LLMs* / *Building LLMs for Production*") # 1. THE LESSON st.markdown("### ๐Ÿ“– The Concept") st.markdown(""" Large Language Models don't just "know" the answer; they predict the next word based on probability. **Temperature** is a hyperparameter that controls how "risky" the model is when choosing that next word. * **Low Temperature (0.0 - 0.3):** The model always picks the most likely word. It is precise, deterministic, and boring. Good for coding. * **High Temperature (0.7 - 1.0+):** The model takes chances on less likely words. It is creative, diverse, but prone to errors (hallucinations). """) # 2. THE SIMULATION st.markdown("### ๐Ÿ•น๏ธ Interactive Simulation") st.info("Adjust the temperature slider below to see how the AI completes the sentence: **'The sky is...'**") temp = st.slider("Temperature", 0.0, 1.5, 0.1) st.markdown('
', unsafe_allow_html=True) st.write(f"SYSTEM: Generating completion with Temperature = {temp}...") # Simulated Logic (Deterministic vs Random) if temp < 0.3: output = "blue." explanation = "(The model picked the probability winner: 99%)" elif temp < 0.7: output = "filled with scattered clouds drifting slowly." explanation = "(The model picked a top-5 probability choice)" elif temp < 1.1: output = "a canvas of infinite azure, weeping rain." explanation = "(The model took a risk on lower probability words like 'weeping')" else: output = "blue fish guitar singing backwards yesterday." explanation = "(Complete chaos. The model picked very low probability tokens)" st.write(f"**OUTPUT:** The sky is {output}") st.markdown("
", unsafe_allow_html=True) st.caption(explanation) # 3. KNOWLEDGE CHECK st.markdown("### โœ… Knowledge Check") q1 = st.radio("Which temperature should you use for a Code Generation assistant?", ["0.1 (Precise)", "0.7 (Balanced)", "1.2 (Creative)"], index=None) if st.button("Submit Answer"): if q1 == "0.1 (Precise)": st.success("Correct! Code requires precision. You don't want 'creative' syntax errors.") if "Mod1" not in st.session_state.completed_modules: st.session_state.xp += 100 st.session_state.completed_modules.add("Mod1") st.balloons() else: st.error("Not quite. High temperature introduces randomness, which breaks code syntax.") # --- MODULE 2: EMBEDDINGS --- elif page == "2. Meaning in Numbers (Embeddings)": st.markdown('
Module 2: How Machines "Understand" Meaning
', unsafe_allow_html=True) st.caption("Reference: *Hands-On Large Language Models* / *Deep Learning*") # 1. THE LESSON st.markdown("### ๐Ÿ“– The Concept") st.markdown(""" Computers cannot understand text. They only understand numbers. **Vector Embeddings** are lists of numbers (coordinates) that represent the *meaning* of a word. Imagine a 2D map: * "King" is close to "Queen". * "Apple" is far away from "King", but close to "Banana". When we want to search for something, we don't match keywords; we measure the **distance** between these points. """) # 2. THE SIMULATION st.markdown("### ๐Ÿ•น๏ธ Vector Laboratory") st.info("See how the math groups related concepts together.") # Mock Data for Visualization data = pd.DataFrame({ 'word': ['King', 'Queen', 'Prince', 'Apple', 'Banana', 'Cherry', 'Car', 'Truck'], 'x': [1, 1.2, 1.1, 5, 5.2, 5.1, 8, 8.2], # X Coordinate (Theme) 'y': [1, 1.1, 0.9, 5, 4.8, 5.2, 1, 1.2], # Y Coordinate (Sub-theme) 'Category': ['Royalty', 'Royalty', 'Royalty', 'Fruit', 'Fruit', 'Fruit', 'Vehicle', 'Vehicle'] }) st.scatter_chart(data, x='x', y='y', color='Category', size=200) st.markdown(""" **Observation:** Notice how the "Royalty" words are clustered in the bottom-left, and "Fruit" is in the top-right? If you search for "Throne", the math looks for points in that bottom-left cluster. """) # 3. KNOWLEDGE CHECK st.markdown("### โœ… Knowledge Check") q2 = st.selectbox("In vector math, 'King' minus 'Man' plus 'Woman' equals...?", ["Select...", "Prince", "Queen", "Apple"]) if st.button("Submit Answer"): if q2 == "Queen": st.success("Correct! This is the most famous example of vector arithmetic preserving semantic relationships.") if "Mod2" not in st.session_state.completed_modules: st.session_state.xp += 100 st.session_state.completed_modules.add("Mod2") st.balloons() else: st.error("Try again. Think about the relationship pairs: Man is to Woman as King is to...?") # --- MODULE 3: RAG ARCHITECTURE --- elif page == "3. The Knowledge Bridge (RAG)": st.markdown('
Module 3: Teaching the AI New Facts
', unsafe_allow_html=True) st.caption("Reference: *AI Engineering* / *Designing Machine Learning Systems*") # 1. THE LESSON st.markdown("### ๐Ÿ“– The Concept") st.markdown(""" Standard LLMs are frozen in time (training cut-off). They don't know your private data or recent news. **RAG (Retrieval Augmented Generation)** fixes this without retraining the model. 1. **Retrieve:** You search your database for relevant text chunks. 2. **Augment:** You paste those chunks into the prompt (context). 3. **Generate:** You tell the LLM "Answer using only the text above." """) # 2. THE SIMULATION st.markdown("### ๐Ÿ•น๏ธ RAG Builder") st.info("Simulate a RAG pipeline query.") user_query = st.text_input("User Query:", value="What is the company vacation policy?") col1, col2 = st.columns(2) with col1: st.markdown("**1. Without RAG**") st.markdown(f""" > **User:** {user_query} > **AI:** I don't know which company you are talking about. I only know public information up to 2023. """) with col2: st.markdown("**2. With RAG**") if st.button("Run Retrieval"): st.markdown("... *Searching Vector Database* ...") st.success("Found 1 Document: 'Employee_Handbook.pdf'") st.code(f""" PROMPT SENT TO AI: Context: "Employees are entitled to 20 days of PTO per year." Question: {user_query} Answer: Based on the handbook, the policy is 20 days per year. """) # 3. KNOWLEDGE CHECK st.markdown("### โœ… Knowledge Check") q3 = st.radio("Why is RAG often better than Fine-Tuning for facts?", ["It makes the model smarter generally.", "It allows you to update facts instantly by changing the database.", "It is faster to train."], index=None) if st.button("Submit Answer"): if "update facts instantly" in str(q3): st.success("Correct! If the vacation policy changes tomorrow, you just update the PDF. You don't need to re-train the brain.") if "Mod3" not in st.session_state.completed_modules: st.session_state.xp += 100 st.session_state.completed_modules.add("Mod3") st.balloons() else: st.error("No. Fine-tuning is better for 'style' or 'behavior'. RAG is better for 'facts' that change.")