Update app.py

app.py

@@ -5,13 +5,14 @@ from dotenv import load_dotenv
 import json
 import textwrap
 import time
-from typing import Any, List
+from typing import Any, List, Optional
 import numpy as np
 from sklearn.metrics.pairwise import cosine_similarity
 import tensorflow as tf
 from tensorflow.keras.models import Sequential
 from tensorflow.keras.layers import Dense, Input
 from tensorflow.keras.utils import to_categorical
+from tensorflow.keras.optimizers import Adam  # Import Adam optimizer
 
 # Load environment variables
 load_dotenv()
@@ -20,7 +21,6 @@ GOOGLE_API_KEY = os.getenv("GOOGLE_API_KEY")
 # Configure Generative AI model
 if GOOGLE_API_KEY:
     genai.configure(api_key=GOOGLE_API_KEY)
-    model = genai.GenerativeModel('gemini-pro')  # You can choose a suitable model.
 else:
     st.error(
         "Google AI Studio API key not found. Please add it to your .env file.  "
@@ -43,17 +43,18 @@ with st.sidebar:
     st.markdown(
         """
         -   **Embeddings**: Numerical representations of text, capturing semantic meaning.
-        -   **Vector Databases**:  Databases optimized for storing and querying vectors.
+        -   **Vector Databases**:  Databases optimized for storing and querying vectors (simulated here).
         -   **Retrieval Augmented Generation (RAG)**:  Combining retrieval with LLM generation.
         -   **Cosine Similarity**:  A measure of similarity between two vectors.
+        -   **Neural Networks**: Using embeddings as input for classification.
         """
     )
     st.subheader("Whitepaper Insights")
     st.markdown(
         """
-    -   Efficient similarity search using vector indexes (e.g., ANN).
-    -   Handling large datasets and scalability.
-    -   Applications of embeddings: search, recommendation, classification.
+        -   Efficient similarity search using vector indexes (e.g., ANN).
+        -   Handling large datasets and scalability considerations.
+        -   Applications of embeddings: search, recommendation, classification, etc.
         """
     )
 
@@ -72,20 +73,40 @@ def display_response(response: Any) -> None:
         st.error("Failed to generate a response.")
 
 
-def generate_embeddings(texts: List[str], model_name: str = 'models/embedding-001') -> List[List[float]]:
-    """Generates embeddings for a list of texts."""
+
+def generate_embeddings(texts: List[str], model_name: str) -> Optional[List[List[float]]]:
+    """Generates embeddings for a list of texts using a specified model.
+
+    Args:
+        texts: List of text strings.
+        model_name: Name of the embedding model.
+
+    Returns:
+        List of embeddings (list of floats) or None on error.
+    """
     try:
         embedding_model = genai.EmbeddingModel(model_name)
         embeddings = embedding_model.embed_content(texts=texts)
-        return [embedding.values for embedding in embeddings.embeddings]  # Extract embedding values
+        return [embedding.values for embedding in embeddings.embeddings]
     except Exception as e:
-        st.error(f"Error generating embeddings: {e}")
-        return []
+        st.error(f"Error generating embeddings with model '{model_name}': {e}")
+        return None
 
 
 
 def generate_with_retry(prompt: str, model_name: str, generation_config: genai.types.GenerationConfig, max_retries: int = 3, delay: int = 5) -> Any:
-    """Generates content with retry logic."""
+    """Generates content with retry logic and error handling.
+
+    Args:
+        prompt: The prompt string.
+        model_name: The name of the language model.
+        generation_config: The generation configuration.
+        max_retries: Maximum number of retries.
+        delay: Delay in seconds between retries.
+
+    Returns:
+        The generated response or None on error.
+    """
     for i in range(max_retries):
         try:
             model = genai.GenerativeModel(model_name)
@@ -98,25 +119,81 @@ def generate_with_retry(prompt: str, model_name: str, generation_config: genai.t
                 st.error(
                     f"Model '{model_name}' is not available or not supported.  Please select a different model."
                 )
-                return None
+                return None  # Return None to signal a non-retryable error
             elif i < max_retries - 1:
                 st.info(f"Retrying in {delay} seconds...")
                 time.sleep(delay)
             else:
-                raise
-    raise Exception("Failed to generate content after maximum retries")
+                st.error(f"Failed to generate content after {max_retries} attempts.  Please check your prompt and model.")
+                return None # Return None after max retries
+    return None  #Should never reach here
+
+
+
+def calculate_similarity(embedding1: List[float], embedding2: List[float]) -> float:
+    """Calculates the cosine similarity between two embeddings."""
+    return cosine_similarity(np.array(embedding1).reshape(1, -1), np.array(embedding2).reshape(1, -1))[0][0]
+
+
+
+def create_and_train_model(
+    embeddings: List[List[float]],
+    labels: List[int],
+    num_classes: int,
+    epochs: int,
+    batch_size: int,
+    learning_rate: float,
+    optimizer_str: str
+) -> tf.keras.Model:
+    """Creates and trains a neural network for classification.
+
+    Args:
+        embeddings: List of input embeddings.
+        labels: List of integer labels.
+        num_classes: Number of classes.
+        epochs: Number of training epochs.
+        batch_size: Batch size for training.
+        learning_rate: Learning rate for the optimizer.
+        optimizer_str: Name of the optimizer ('adam', 'sgd', 'rmsprop')
+
+    Returns:
+        Trained Keras model.
+    """
+    model = Sequential([
+        Input(shape=(len(embeddings[0]),)),
+        Dense(64, activation='relu'),  # Increased hidden layer size
+        Dense(32, activation='relu'),
+        Dense(num_classes, activation='softmax')
+    ])
+
+    if optimizer_str.lower() == 'adam':
+        optimizer = Adam(learning_rate=learning_rate)
+    elif optimizer_str.lower() == 'sgd':
+        optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate)
+    elif optimizer_str.lower() == 'rmsprop':
+        optimizer = tf.keras.optimizers.RMSprop(learning_rate=learning_rate)
+    else:
+        optimizer = Adam(learning_rate=learning_rate) #default
+
+    model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['accuracy'])
+    encoded_labels = to_categorical(labels, num_classes=num_classes)
+    model.fit(np.array(embeddings), encoded_labels, epochs=epochs, batch_size=batch_size, verbose=0)
+    return model
 
 
 
 # --- RAG Question Answering ---
 st.header("RAG Question Answering")
 rag_model_name = st.selectbox("Select model for RAG:", ["gemini-pro"], index=0)
+rag_embedding_model = st.selectbox("Select embedding model for RAG:", ["models/embedding-001"], index=0)
 rag_context = st.text_area(
     "Enter your context documents:",
     "Relevant information to answer the question.  Separate documents with newlines.",
     height=150,
 )
-rag_question = st.text_area("Ask a question about the context:", "What is the main topic?", height=50)
+rag_question = st.text_area("Ask a question about the context:", "What is the main topic?", height=70) # Changed height to 70
+rag_max_context_length = st.number_input("Maximum Context Length", min_value=100, max_value=2000, value=500, step=100)
+
 
 if st.button("Answer with RAG"):
     if not rag_context or not rag_question:
@@ -125,12 +202,14 @@ if st.button("Answer with RAG"):
         with st.spinner("Generating answer..."):
             try:
                 # 1. Generate embeddings for the context
-                context_embeddings = generate_embeddings(rag_context.split('\n'))
+                context_embeddings = generate_embeddings(rag_context.split('\n'), rag_embedding_model)
                 if not context_embeddings:
                     st.stop()
 
                 # 2. Generate embedding for the question
-                question_embedding = generate_embeddings([rag_question])[0]
+                question_embedding = generate_embeddings([rag_question], rag_embedding_model)
+                if not question_embedding:
+                    st.stop()
 
                 # 3. Calculate similarity scores
                 similarities = cosine_similarity(np.array(question_embedding).reshape(1, -1), np.array(context_embeddings))[0]
@@ -138,13 +217,17 @@ if st.button("Answer with RAG"):
                 # 4. Find the most relevant document(s)
                 most_relevant_index = np.argmax(similarities)
                 relevant_context = rag_context.split('\n')[most_relevant_index]
+                #truncate context
+                if len(relevant_context) > rag_max_context_length:
+                    relevant_context = relevant_context[:rag_max_context_length]
 
                 # 5. Construct the prompt
                 rag_prompt = f"Use the following context to answer the question: '{rag_question}'.\nContext: {relevant_context}"
 
                 # 6. Generate the answer
                 response = generate_with_retry(rag_prompt, rag_model_name, generation_config=genai.types.GenerationConfig())
-                display_response(response)
+                if response:
+                    display_response(response)
             except Exception as e:
                 st.error(f"An error occurred: {e}")
 
@@ -152,9 +235,9 @@ if st.button("Answer with RAG"):
 
 # --- Text Similarity ---
 st.header("Text Similarity")
-similarity_model_name = st.selectbox("Select model for similarity:", ["models/embedding-001"], index=0)  # Use a model that supports embeddings
-text1 = st.text_area("Enter text 1:", "This is the first sentence.", height=50)
-text2 = st.text_area("Enter text 2:", "This is a similar sentence.", height=50)
+similarity_embedding_model = st.selectbox("Select embedding model for similarity:", ["models/embedding-001"], index=0)
+text1 = st.text_area("Enter text 1:", "This is the first sentence.", height=70) # Changed height to 70
+text2 = st.text_area("Enter text 2:", "This is a similar sentence.", height=70) # Changed height to 70
 
 if st.button("Calculate Similarity"):
     if not text1 or not text2:
@@ -163,11 +246,12 @@ if st.button("Calculate Similarity"):
         with st.spinner("Calculating similarity..."):
             try:
                 # 1. Generate embeddings
-                embeddings = generate_embeddings([text1, text2], similarity_model_name)
+                embeddings = generate_embeddings([text1, text2], similarity_embedding_model)
                 if not embeddings:
                     st.stop()
+
                 # 2. Calculate cosine similarity
-                similarity = cosine_similarity([embeddings[0]], [embeddings[1]])[0][0]
+                similarity = calculate_similarity(embeddings[0], embeddings[1])
                 st.subheader("Cosine Similarity:")
                 st.write(similarity)
             except Exception as e:
@@ -177,28 +261,48 @@ if st.button("Calculate Similarity"):
 
 # --- Neural Classification ---
 st.header("Neural Classification with Embeddings")
-classification_model_name = st.selectbox("Select model for classification:", ["models/embedding-001"], index=0) #use embedding model
+classification_embedding_model = st.selectbox("Select embedding model for classification:", ["models/embedding-001"], index=0)
 classification_data = st.text_area(
     "Enter your training data (text, label pairs), separated by newlines.  Example: text1,0\\ntext2,1",
     "text1,0\ntext2,1\ntext3,0\ntext4,1",
     height=150,
 )
-classification_prompt = st.text_area("Enter text to classify:", "This is a test text.", height=50)
-num_epochs = st.number_input("Number of Epochs", min_value=1, max_value=100, value=10, step=1)
-
-
-def create_and_train_model(embeddings: List[List[float]], labels: List[int], num_classes: int, epochs: int):
-    """Creates and trains a simple neural network for classification."""
-    model = Sequential([
-        Input(shape=(len(embeddings[0]),)),  # Input shape is the embedding size
-        Dense(16, activation='relu'),
-        Dense(num_classes, activation='softmax')  # Output layer with softmax
-    ])
-
-    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
-    encoded_labels = to_categorical(labels, num_classes=num_classes) #one hot encode
-    model.fit(np.array(embeddings), encoded_labels, epochs=epochs, verbose=0)  # Suppress training output
-    return model
+classification_prompt = st.text_area("Enter text to classify:", "This is a test text.", height=70) # Changed height to 70
+num_epochs = st.number_input("Number of Epochs", min_value=1, max_value=200, value=10, step=1)
+batch_size = st.number_input("Batch Size", min_value=1, max_value=128, value=32, step=1)
+learning_rate = st.number_input("Learning Rate", min_value=0.0001, max_value=0.1, value=0.0001, step=0.0001, format="%.4f")
+optimizer_str = st.selectbox("Optimizer", ['adam', 'sgd', 'rmsprop'], index=0)
+
+
+def process_classification_data(data: str) -> Optional[tuple[List[str], List[int]]]:
+    """Processes the classification data string into lists of texts and labels.
+
+    Args:
+        data:  The input string.
+
+    Returns:
+       A tuple of (texts, labels) or None on error
+    """
+    data_pairs = [line.split(',') for line in data.split('\n') if ',' in line]
+    if not data_pairs:
+        st.error("No valid data pairs found.  Please ensure each line contains 'text,label'.")
+        return None
+    texts = []
+    labels = []
+    for i, pair in enumerate(data_pairs):
+        if len(pair) != 2:
+            st.error(f"Invalid data format in line {i + 1}: '{','.join(pair)}'.  Expected 'text,label'.")
+            return None
+        text = pair[0].strip()
+        label_str = pair[1].strip()
+        try:
+            label = int(label_str)
+            texts.append(text)
+            labels.append(label)
+        except ValueError:
+            st.error(f"Invalid label value in line {i + 1}: '{label_str}'.  Label must be an integer.")
+            return None
+    return texts, labels
 
 
 
@@ -209,21 +313,26 @@ if st.button("Classify"):
         with st.spinner("Classifying..."):
             try:
                 # 1. Process the training data
-                data_pairs = [line.split(',') for line in classification_data.split('\n') if ',' in line]
-                train_texts = [pair[0].strip() for pair in data_pairs]
-                train_labels = [int(pair[1].strip()) for pair in data_pairs]
-                num_classes = len(set(train_labels)) #number of classes
+                processed_data = process_classification_data(classification_data)
+                if not processed_data:
+                    st.stop()
+                train_texts, train_labels = processed_data
+                num_classes = len(set(train_labels))
 
                 # 2. Generate embeddings for training data
-                train_embeddings = generate_embeddings(train_texts, classification_model_name)
+                train_embeddings = generate_embeddings(train_texts, classification_embedding_model)
                 if not train_embeddings:
                     st.stop()
 
                 # 3. Create and train the model
-                model = create_and_train_model(train_embeddings, train_labels, num_classes, num_epochs)
+                model = create_and_train_model(
+                    train_embeddings, train_labels, num_classes, num_epochs, batch_size, learning_rate, optimizer_str
+                )
 
                 # 4. Generate embedding for the text to classify
-                predict_embedding = generate_embeddings([classification_prompt], classification_model_name)[0]
+                predict_embedding = generate_embeddings([classification_prompt], classification_embedding_model)
+                if not predict_embedding:
+                    st.stop()
 
                 # 5. Make the prediction
                 prediction = model.predict(np.array([predict_embedding]), verbose=0)
@@ -235,4 +344,3 @@ if st.button("Classify"):
 
             except Exception as e:
                 st.error(f"An error occurred: {e}")
-