import streamlit as st
import pandas as pd
import plotly.express as px
import numpy as np
from datetime import datetime, timedelta
import json
from pymongo import MongoClient
from dotenv import load_dotenv
import os
import bcrypt
from openai import OpenAI
from streamlit_plotly_events import plotly_events
from pinecone import Pinecone, ServerlessSpec
import threading # {{ edit_25: Import threading for background processing }}
import tiktoken
from tiktoken.core import Encoding
from runner import run_model
from bson.objectid import ObjectId
import traceback # Add this import at the top of your file
import umap
import plotly.graph_objs as go
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans
import plotly.colors as plc
# Add this helper function at the beginning of your file
def extract_prompt_text(prompt):
if isinstance(prompt, dict):
return prompt.get('prompt', '')
elif isinstance(prompt, str):
return prompt
else:
return str(prompt)
# Set page configuration to wide mode
st.set_page_config(layout="wide")
# Load environment variables
load_dotenv()
# MongoDB connection
mongodb_uri = os.getenv('MONGODB_URI')
mongo_client = MongoClient(mongodb_uri) # {{ edit_11: Rename MongoDB client to 'mongo_client' }}
db = mongo_client['llm_evaluation_system']
users_collection = db['users']
results_collection = db['evaluation_results']
# Remove or comment out this line if it exists
# openai_client = OpenAI()
# Instead, use the openai_client from runner.py
from runner import openai_client
# Initialize Pinecone
pinecone_client = Pinecone(api_key=os.getenv('PINECONE_API_KEY')) # {{ edit_13: Initialize Pinecone client using Pinecone class }}
# Initialize the tokenizer
tokenizer: Encoding = tiktoken.get_encoding("cl100k_base") # This is suitable for GPT-4 and recent models
# Authentication functions
def hash_password(password):
return bcrypt.hashpw(password.encode('utf-8'), bcrypt.gensalt())
def verify_password(password, hashed_password):
return bcrypt.checkpw(password.encode('utf-8'), hashed_password)
def authenticate(username, password):
user = users_collection.find_one({"username": username})
if user and verify_password(password, user['password']):
return True
return False
def signup(username, password):
if users_collection.find_one({"username": username}):
return False
hashed_password = hash_password(password)
# {{ edit_1: Initialize models list for the new user }}
users_collection.insert_one({
"username": username,
"password": hashed_password,
"models": [] # List to store user's models
})
return True
def upload_model(file):
return "Model uploaded successfully!"
# Function to perform evaluation (placeholder)
def evaluate_model(model_identifier, metrics, username):
# {{ edit_4: Differentiate between Custom and Named models }}
user = users_collection.find_one({"username": username})
models = user.get("models", [])
selected_model = next((m for m in models if (m['model_name'] == model_identifier) or (m['model_id'] == model_identifier)), None)
if selected_model:
if selected_model.get("model_type") == "named":
# For Named Models, use RAG-based evaluation
return evaluate_named_model(model_identifier, prompt, context_dataset)
else:
# For Custom Models, proceed with existing evaluation logic
results = {metric: round(np.random.rand() * 100, 2) for metric in metrics}
return results
else:
st.error("Selected model not found.")
return None
# Function to generate response using GPT-4-mini
def generate_response(prompt, context):
try:
response = openai_client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": f"Context: {context}\n\nPrompt: {prompt}"}
]
)
return response.choices[0].message.content
except Exception as e:
st.error(f"Error generating response: {str(e)}")
return None
# Add this function to update the context for a model
def update_model_context(username, model_id, context):
users_collection.update_one(
{"username": username, "models.model_id": model_id},
{"$set": {"models.$.context": context}}
)
# Function to clear the results database
def clear_results_database(username, model_identifier=None):
try:
if model_identifier:
# Clear results for the specific model
results_collection.delete_many({
"username": username,
"$or": [
{"model_name": model_identifier},
{"model_id": model_identifier}
]
})
else:
# Clear all results for the user
results_collection.delete_many({"username": username})
return True
except Exception as e:
st.error(f"Error clearing results database: {str(e)}")
return False
# Function to generate embeddings using the specified model
def generate_embedding(text):
try:
embedding_response = openai_client.embeddings.create(
model="text-embedding-3-large", # {{ edit_3: Use the specified embedding model }}
input=text,
encoding_format="float"
)
embedding = embedding_response["data"][0]["embedding"]
return embedding
except Exception as e:
st.error(f"Error generating embedding: {str(e)}")
return None
# Function to handle Named Model Evaluation using RAG
def evaluate_named_model(model_name, prompt, context_dataset):
# {{ edit_4: Implement evaluation using RAG and Pinecone with the specified embedding model }}
try:
# Initialize Pinecone index
index = pinecone_client.Index(os.getenv('PINECONE_INDEX_NAME'))
# Generate embedding for the prompt
prompt_embedding = generate_embedding(prompt)
if not prompt_embedding:
st.error("Failed to generate embedding for the prompt.")
return None
# Retrieve relevant context using RAG by querying Pinecone with the embedding
query_response = index.query(
top_k=5,
namespace=model_name,
include_metadata=True,
vector=prompt_embedding # {{ edit_5: Use embedding vector for querying }}
)
# Aggregate retrieved context
retrieved_context = " ".join([item['metadata']['text'] for item in query_response['matches']])
# Generate response using the retrieved context
response = generate_response(prompt, retrieved_context)
# Evaluate the response
evaluation = teacher_evaluate(prompt, retrieved_context, response)
# Save the results
save_results(model_name, prompt, retrieved_context, response, evaluation)
return evaluation
except Exception as e:
st.error(f"Error in evaluating named model: {str(e)}")
return None
# Example: When indexing data to Pinecone, generate embeddings using the specified model
def index_context_data(model_name, texts):
try:
index = pinecone_client.Index(os.getenv('PINECONE_INDEX_NAME'))
for text in texts:
embedding = generate_embedding(text)
if embedding:
index.upsert([
{
"id": f"{model_name}_{hash(text)}",
"values": embedding,
"metadata": {"text": text}
}
])
except Exception as e:
st.error(f"Error indexing data to Pinecone: {str(e)}")
def upload_model(file, username, model_type):
# {{ edit_5: Modify upload_model to handle model_type }}
model_id = f"{username}_model_{int(datetime.now().timestamp())}"
if model_type == "custom":
# Save the model file as needed
model_path = os.path.join("models", f"{model_id}.bin")
with open(model_path, "wb") as f:
f.write(file.getbuffer())
# Update user's models list
users_collection.update_one(
{"username": username},
{"$push": {"models": {
"model_id": model_id,
"file_path": model_path,
"uploaded_at": datetime.now(),
"model_type": "custom"
}}}
)
return f"Custom Model {model_id} uploaded successfully!"
elif model_type == "named":
# For Named Models, assume the model is managed externally (e.g., via Pinecone)
users_collection.update_one(
{"username": username},
{"$push": {"models": {
"model_id": model_id,
"model_name": None,
"file_path": None,
"model_link": None,
"uploaded_at": datetime.now(),
"model_type": "named"
}}}
)
return f"Named Model {model_id} registered successfully!"
else:
return "Invalid model type specified."
# Function to save results to MongoDB
def save_results(username, model, prompt, context, response, evaluation): # {{ edit_29: Add 'username' parameter }}
result = {
"username": username, # Use the passed 'username' parameter
"model_id": model['model_id'], # {{ edit_19: Associate results with 'model_id' }}
"model_name": model.get('model_name'),
"model_type": model.get('model_type', 'custom'), # {{ edit_20: Include 'model_type' in results }}
"prompt": prompt,
"context": context,
"response": response,
"evaluation": evaluation,
"timestamp": datetime.now()
}
results_collection.insert_one(result)
# Modify the run_custom_evaluations function
def run_custom_evaluations(data, selected_model, username):
try:
model_name = selected_model['model_name']
model_id = selected_model['model_id']
model_type = selected_model.get('model_type', 'Unknown').lower()
if model_type == 'simple':
# For simple models, data is already in the correct format
test_cases = data
else:
# For other models, data is split into context_dataset and questions
context_dataset, questions = data
test_cases = [
{
"prompt": extract_prompt_text(question),
"context": context_dataset,
"response": "" # This will be filled by the model
}
for question in questions
]
for test_case in test_cases:
prompt_text = test_case["prompt"]
context = test_case["context"]
# Get the student model's response using runner.py
try:
answer = run_model(model_name, prompt_text)
if answer is None or answer == "":
st.warning(f"No response received from the model for prompt: {prompt_text}")
answer = "No response received from the model."
except Exception as model_error:
st.error(f"Error running model for prompt: {prompt_text}")
st.error(f"Error details: {str(model_error)}")
answer = f"Error: {str(model_error)}"
# Get the teacher's evaluation
try:
evaluation = teacher_evaluate(prompt_text, context, answer)
if evaluation is None:
st.warning(f"No evaluation received for prompt: {prompt_text}")
evaluation = {"Error": "No evaluation received"}
except Exception as eval_error:
st.error(f"Error in teacher evaluation for prompt: {prompt_text}")
st.error(f"Error details: {str(eval_error)}")
evaluation = {"Error": str(eval_error)}
# Save the results
save_results(username, selected_model, prompt_text, context, answer, evaluation)
st.success("Evaluation completed successfully!")
except Exception as e:
st.error(f"Error in custom evaluation: {str(e)}")
st.error(f"Detailed error: {traceback.format_exc()}")
# Function for teacher model evaluation
def teacher_evaluate(prompt, context, response):
try:
evaluation_prompt = f"""
Evaluate the following response based on the given prompt and context.
Rate each factor on a scale of 0 to 1, where 1 is the best (or least problematic for negative factors like Hallucination and Bias).
Please provide scores with two decimal places, and avoid extreme scores of exactly 0 or 1 unless absolutely necessary.
Context: {context}
Prompt: {prompt}
Response: {response}
Factors to evaluate:
1. Accuracy: How factually correct is the response?
2. Hallucination: To what extent does the response contain made-up information? (Higher score means less hallucination)
3. Groundedness: How well is the response grounded in the given context and prompt?
4. Relevance: How relevant is the response to the prompt?
5. Recall: How much of the relevant information from the context is included in the response?
6. Precision: How precise and focused is the response in addressing the prompt?
7. Consistency: How consistent is the response with the given information and within itself?
8. Bias Detection: To what extent is the response free from bias? (Higher score means less bias)
Provide the evaluation as a JSON object. Each factor should be a key mapping to an object containing 'score' and 'explanation'.
Do not include any additional text, explanations, or markdown formatting.
"""
evaluation_response = openai_client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": "You are an expert evaluator of language model responses."},
{"role": "user", "content": evaluation_prompt}
]
)
content = evaluation_response.choices[0].message.content.strip()
# Ensure the response starts and ends with curly braces
if not (content.startswith("{") and content.endswith("}")):
st.error("Teacher evaluation did not return a valid JSON object.")
st.error(f"Response content: {content}")
return None
try:
evaluation = json.loads(content)
return evaluation
except json.JSONDecodeError as e:
st.error(f"Error decoding evaluation response: {str(e)}")
st.error(f"Response content: {content}")
return None
except Exception as e:
st.error(f"Error in teacher evaluation: {str(e)}")
return None
# Function to generate dummy data for demonstration
def generate_dummy_data():
dates = pd.date_range(end=datetime.now(), periods=30).tolist()
metrics = ['Accuracy', 'Precision', 'Recall', 'F1 Score', 'Consistency', 'Bias']
data = {
'Date': dates * len(metrics),
'Metric': [metric for metric in metrics for _ in range(len(dates))],
'Value': np.random.rand(len(dates) * len(metrics)) * 100
}
return pd.DataFrame(data)
# Function to count tokens
def count_tokens(text: str) -> int:
return len(tokenizer.encode(text))
# Sidebar Navigation
st.sidebar.title("LLM Evaluation System")
# Session state
if 'user' not in st.session_state:
st.session_state.user = None
# Authentication
if not st.session_state.user:
auth_option = st.sidebar.radio("Choose an option", ["Login", "Signup"])
username = st.sidebar.text_input("Username")
password = st.sidebar.text_input("Password", type="password")
if auth_option == "Login":
if st.sidebar.button("Login"):
if authenticate(username, password):
st.session_state.user = username
st.rerun()
else:
st.sidebar.error("Invalid username or password")
else:
if st.sidebar.button("Signup"):
if signup(username, password):
st.sidebar.success("Signup successful. Please login.")
else:
st.sidebar.error("Username already exists")
else:
st.sidebar.success(f"Welcome, {st.session_state.user}!")
if st.sidebar.button("Logout"):
st.session_state.user = None
st.rerun()
# App content
if st.session_state.user:
app_mode = st.sidebar.selectbox("Choose the section", ["Dashboard", "Model Upload", "Evaluation", "Prompt Testing", "Manage Models", "History"]) # {{ edit_add: Added "History" to the sidebar navigation }}
if app_mode == "Dashboard":
st.title("Dashboard")
st.write("### Real-time Metrics and Performance Insights")
# Fetch the user from the database
user = users_collection.find_one({"username": st.session_state.user})
if user is None:
st.error("User not found in the database.")
st.stop()
user_models = user.get("models", [])
if user_models:
model_options = [model['model_name'] if model['model_name'] else model['model_id'] for model in user_models]
selected_model = st.selectbox("Select Model to View Metrics", ["All Models"] + model_options)
st.session_state['selected_model'] = selected_model # Store the selected model in session state
# Add delete dataset button
if selected_model != "All Models":
if st.button("Delete Dataset"):
if st.session_state['selected_model']:
if clear_results_database(st.session_state.user, st.session_state['selected_model']):
st.success(f"All evaluation results for {st.session_state['selected_model']} have been deleted.")
st.rerun() # Rerun the app to refresh the dashboard
else:
st.error("Failed to delete the dataset. Please try again.")
else:
st.error("No model selected. Please select a model to delete its dataset.")
else:
st.error("You have no uploaded models.")
selected_model = "All Models"
st.session_state['selected_model'] = selected_model
try:
query = {"username": st.session_state.user}
if selected_model != "All Models":
query["model_name"] = selected_model
if not selected_model:
query = {"username": st.session_state.user, "model_id": selected_model}
results = list(results_collection.find(query))
if results:
df = pd.DataFrame(results)
# Check if required columns exist
required_columns = ['prompt', 'context', 'response', 'evaluation']
missing_columns = [col for col in required_columns if col not in df.columns]
if missing_columns:
st.error(f"Error: Missing columns in the data: {', '.join(missing_columns)}")
st.error("Please check the database schema and ensure all required fields are present.")
st.stop()
# Extract prompt text if needed
df['prompt'] = df['prompt'].apply(extract_prompt_text)
# Safely count tokens for prompt, context, and response
def safe_count_tokens(text):
if isinstance(text, str):
return count_tokens(text)
else:
return 0 # or some default value
df['prompt_tokens'] = df['prompt'].apply(safe_count_tokens)
df['context_tokens'] = df['context'].apply(safe_count_tokens)
df['response_tokens'] = df['response'].apply(safe_count_tokens)
# Calculate total tokens for each row
df['total_tokens'] = df['prompt_tokens'] + df['context_tokens'] + df['response_tokens']
# Safely extract evaluation metrics
metrics = ["Accuracy", "Hallucination", "Groundedness", "Relevance", "Recall", "Precision", "Consistency", "Bias Detection"]
for metric in metrics:
df[metric] = df['evaluation'].apply(lambda x: x.get(metric, {}).get('score', 0) if isinstance(x, dict) else 0) * 100
df['timestamp'] = pd.to_datetime(df['timestamp'])
df['query_number'] = range(1, len(df) + 1) # Add query numbers
# Set the threshold for notifications
notification_threshold = st.slider("Set Performance Threshold for Notifications (%)", min_value=0, max_value=100, value=50)
# Define the metrics to check
metrics_to_check = metrics # Or allow the user to select specific metrics
# Check for evaluations where any of the metrics are below the threshold
low_performance_mask = df[metrics_to_check].lt(notification_threshold).any(axis=1)
low_performing_evaluations = df[low_performance_mask]
# Display Notifications
if not low_performing_evaluations.empty:
st.warning(f"⚠️ You have {len(low_performing_evaluations)} evaluations with metrics below {notification_threshold}%.")
with st.expander("View Low-Performing Evaluations"):
# Display the low-performing evaluations in a table
display_columns = ['timestamp', 'model_name', 'prompt', 'response'] + metrics_to_check
low_perf_display_df = low_performing_evaluations[display_columns].copy()
low_perf_display_df['timestamp'] = low_perf_display_df['timestamp'].dt.strftime('%Y-%m-%d %H:%M:%S')
# Apply styling to highlight low scores
def highlight_low_scores(val):
if isinstance(val, float):
if val < notification_threshold:
return 'background-color: red; color: white'
return ''
styled_low_perf_df = low_perf_display_df.style.applymap(highlight_low_scores, subset=metrics_to_check)
styled_low_perf_df = styled_low_perf_df.format({metric: "{:.2f}%" for metric in metrics_to_check})
st.dataframe(
styled_low_perf_df.set_properties(**{
'text-align': 'left',
'border': '1px solid #ddd'
}).set_table_styles([
{'selector': 'th', 'props': [('background-color', '#333'), ('color', 'white')]},
{'selector': 'td', 'props': [('vertical-align', 'top')]}
]),
use_container_width=True
)
else:
st.success("🎉 All your evaluations have metrics above the threshold!")
@st.cache_data
def create_metrics_graph(df, metrics):
fig = px.line(
df,
x='query_number', # Use query numbers on x-axis
y=metrics,
title='Metrics Over Queries',
labels={metric: f"{metric} (%)" for metric in metrics},
markers=True,
template='plotly_dark',
)
color_discrete_sequence = px.colors.qualitative.Dark24
for i, metric in enumerate(metrics):
fig.data[i].line.color = color_discrete_sequence[i % len(color_discrete_sequence)]
fig.data[i].marker.color = color_discrete_sequence[i % len(color_discrete_sequence)]
fig.update_layout(
xaxis_title="Query Number",
yaxis_title="Metric Score (%)",
legend_title="Metrics",
hovermode="x unified",
margin=dict(l=50, r=50, t=100, b=50),
height=700 # Increase the height of the graph
)
return fig
fig = create_metrics_graph(df, metrics)
st.plotly_chart(fig, use_container_width=True)
# Latest Metrics
st.subheader("Latest Metrics")
latest_metrics = df[metrics].mean() # Calculate the average of all metrics
cols = st.columns(4)
for i, (metric, value) in enumerate(latest_metrics.items()):
with cols[i % 4]:
color = 'green' if value >= 75 else 'orange' if value >= 50 else 'red'
st.metric(label=metric, value=f"{value:.2f}%", delta=None)
st.progress(value / 100)
# Add an explanation for the metrics
st.info("These metrics represent the average scores across all evaluations.")
# Detailed Data View
st.subheader("Detailed Data View")
# Calculate aggregate metrics
total_spans = len(df)
total_tokens = df['total_tokens'].sum()
# Display aggregate metrics
col1, col2 = st.columns(2)
with col1:
st.metric("Total Spans", f"{total_spans:,}")
with col2:
st.metric("Total Tokens", f"{total_tokens:,.2f}M" if total_tokens >= 1e6 else f"{total_tokens:,}")
# Prepare the data for display
display_data = []
for _, row in df.iterrows():
prompt_text = extract_prompt_text(row.get('prompt', ''))
display_row = {
"Prompt": prompt_text[:50] + "..." if prompt_text else "N/A",
"Context": str(row.get('context', ''))[:50] + "..." if row.get('context') else "N/A",
"Response": str(row.get('response', ''))[:50] + "..." if row.get('response') else "N/A",
}
# Add metrics to the display row
for metric in metrics:
display_row[metric] = row.get(metric, 0) # Use get() with a default value
display_data.append(display_row)
# Convert to DataFrame for easy display
display_df = pd.DataFrame(display_data)
# Function to color cells based on score
def color_cells(val):
if isinstance(val, float):
if val >= 80:
color = 'green'
elif val >= 60:
color = '#90EE90' # Light green
else:
color = 'red'
return f'background-color: {color}; color: black'
return ''
# Apply the styling only to metric columns
styled_df = display_df.style.applymap(color_cells, subset=metrics)
# Format metric columns as percentages
for metric in metrics:
styled_df = styled_df.format({metric: "{:.2f}%"})
# Display the table with custom styling
st.dataframe(
styled_df.set_properties(**{
'color': 'white',
'border': '1px solid #ddd'
}).set_table_styles([
{'selector': 'th', 'props': [('background-color', '#4CAF50'), ('color', 'white')]},
{'selector': 'td', 'props': [('text-align', 'left')]},
# Keep background white for non-metric columns
{'selector': 'td:nth-child(-n+3)', 'props': [('background-color', 'white !important')]}
]),
use_container_width=True,
height=400 # Set a fixed height with scrolling
)
# UMAP Visualization with Clustering
st.subheader("UMAP Visualization with Clustering")
if len(df) > 2:
# Allow user to select metrics to include
metrics = ['Accuracy', 'Hallucination', 'Groundedness', 'Relevance', 'Recall', 'Precision', 'Consistency', 'Bias Detection']
selected_metrics = st.multiselect("Select Metrics to Include in UMAP", metrics, default=metrics)
if len(selected_metrics) < 2:
st.warning("Please select at least two metrics for UMAP.")
else:
# Allow user to select number of dimensions
n_components = st.radio("Select UMAP Dimensions", [2, 3], index=1)
# Allow user to adjust UMAP parameters
n_neighbors = st.slider("n_neighbors", min_value=2, max_value=50, value=15)
min_dist = st.slider("min_dist", min_value=0.0, max_value=1.0, value=0.1, step=0.01)
# Prepare data for UMAP
X = df[selected_metrics].values
# Normalize the data
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
# Perform UMAP dimensionality reduction
reducer = umap.UMAP(n_neighbors=n_neighbors, min_dist=min_dist, n_components=n_components, random_state=42)
embedding = reducer.fit_transform(X_scaled)
# Allow user to select the number of clusters
num_clusters = st.slider("Select Number of Clusters", min_value=2, max_value=10, value=3)
# Perform KMeans clustering on the UMAP embeddings
kmeans = KMeans(n_clusters=num_clusters, random_state=42)
cluster_labels = kmeans.fit_predict(embedding)
# Create a DataFrame with the UMAP results and cluster labels
umap_columns = [f'UMAP{i+1}' for i in range(n_components)]
umap_data = {col: embedding[:, idx] for idx, col in enumerate(umap_columns)}
umap_data['Cluster'] = cluster_labels
umap_data['Model'] = df['model_name']
umap_data['Prompt'] = df['prompt']
umap_data['Response'] = df['response']
umap_data['Timestamp'] = df['timestamp']
umap_df = pd.DataFrame(umap_data)
# Include selected metrics in umap_df for hover info
for metric in selected_metrics:
umap_df[metric] = df[metric]
# Prepare customdata for hovertemplate
customdata_columns = ['Model', 'Prompt', 'Cluster'] + selected_metrics
umap_df['customdata'] = umap_df[customdata_columns].values.tolist()
# Build hovertemplate
hovertemplate = 'Model: %{customdata[0]}
' + \
'Prompt: %{customdata[1]}
' + \
'Cluster: %{customdata[2]}
'
for idx, metric in enumerate(selected_metrics):
hovertemplate += f'{metric}: %{{customdata[{idx+3}]:.2f}}
'
hovertemplate += '' # Hide trace info
# Define color palette for clusters
cluster_colors = plc.qualitative.Plotly
num_colors = len(cluster_colors)
if num_clusters > num_colors:
cluster_colors = plc.sample_colorscale('Rainbow', [n/(num_clusters-1) for n in range(num_clusters)])
else:
cluster_colors = cluster_colors[:num_clusters]
# Map cluster labels to colors
cluster_color_map = {label: color for label, color in zip(range(num_clusters), cluster_colors)}
umap_df['Color'] = umap_df['Cluster'].map(cluster_color_map)
# Create the UMAP plot
if n_components == 3:
# 3D plot
fig = go.Figure()
for cluster_label in sorted(umap_df['Cluster'].unique()):
cluster_data = umap_df[umap_df['Cluster'] == cluster_label]
fig.add_trace(go.Scatter3d(
x=cluster_data['UMAP1'],
y=cluster_data['UMAP2'],
z=cluster_data['UMAP3'],
mode='markers',
name=f'Cluster {cluster_label}',
marker=dict(
size=5,
color=cluster_data['Color'], # Color according to cluster
opacity=0.8,
line=dict(width=0.5, color='white')
),
customdata=cluster_data['customdata'],
hovertemplate=hovertemplate
))
fig.update_layout(
title='3D UMAP Visualization with Clustering',
scene=dict(
xaxis_title='UMAP Dimension 1',
yaxis_title='UMAP Dimension 2',
zaxis_title='UMAP Dimension 3'
),
hovermode='closest',
template='plotly_dark',
height=800,
legend_title='Clusters'
)
st.plotly_chart(fig, use_container_width=True)
else:
# 2D plot
fig = go.Figure()
for cluster_label in sorted(umap_df['Cluster'].unique()):
cluster_data = umap_df[umap_df['Cluster'] == cluster_label]
fig.add_trace(go.Scatter(
x=cluster_data['UMAP1'],
y=cluster_data['UMAP2'],
mode='markers',
name=f'Cluster {cluster_label}',
marker=dict(
size=8,
color=cluster_data['Color'], # Color according to cluster
opacity=0.8,
line=dict(width=0.5, color='white')
),
customdata=cluster_data['customdata'],
hovertemplate=hovertemplate
))
fig.update_layout(
title='2D UMAP Visualization with Clustering',
xaxis_title='UMAP Dimension 1',
yaxis_title='UMAP Dimension 2',
hovermode='closest',
template='plotly_dark',
height=800,
legend_title='Clusters'
)
st.plotly_chart(fig, use_container_width=True)
# Selectable Data Points
st.subheader("Cluster Analysis")
# Show cluster counts
cluster_counts = umap_df['Cluster'].value_counts().sort_index().reset_index()
cluster_counts.columns = ['Cluster', 'Number of Points']
st.write("### Cluster Summary")
st.dataframe(cluster_counts)
# Allow user to select clusters to view details
selected_clusters = st.multiselect("Select Clusters to View Details", options=sorted(umap_df['Cluster'].unique()), default=sorted(umap_df['Cluster'].unique()))
if selected_clusters:
selected_data = umap_df[umap_df['Cluster'].isin(selected_clusters)]
st.write("### Details of Selected Clusters")
st.dataframe(selected_data[['Model', 'Prompt', 'Response', 'Cluster'] + selected_metrics])
else:
st.info("Select clusters to view their details.")
st.info("""
**UMAP Visualization with Clustering**
This visualization includes clustering of the evaluation data points in the UMAP space.
**Features:**
- **Clustering Algorithm**: KMeans clustering is applied on the UMAP embeddings.
- **Cluster Selection**: Choose the number of clusters to identify patterns in the data.
- **Color Coding**: Each cluster is represented by a distinct color in the plot.
- **Interactive Exploration**: Hover over points to see detailed information, including the cluster label.
- **Cluster Analysis**: View summary statistics and details of selected clusters.
**Instructions:**
- **Select Metrics**: Choose which evaluation metrics to include in the UMAP calculation.
- **Adjust UMAP Parameters**: Fine-tune `n_neighbors` and `min_dist` for clustering granularity.
- **Choose Number of Clusters**: Use the slider to set how many clusters to identify.
- **Interact with the Plot**: Hover and click on clusters to explore data points.
**Interpreting Clusters:**
- **Cluster Composition**: Clusters group evaluations with similar metric profiles.
- **Model Performance**: Analyze clusters to identify strengths and weaknesses of models.
- **Data Patterns**: Use clustering to uncover hidden patterns in your evaluation data.
**Tips:**
- Experiment with different numbers of clusters to find meaningful groupings.
- Adjust UMAP parameters to see how the clustering changes with different embeddings.
- Use the cluster details to investigate specific evaluations and prompts.
Enjoy exploring your evaluation data with clustering!
""")
else:
st.info("Not enough data for UMAP visualization. Please run more evaluations.")
# Worst Performing Slice Analysis
st.subheader("Worst Performing Slice Analysis")
# Allow the user to select metrics to analyze
metrics = ['Accuracy', 'Hallucination', 'Groundedness', 'Relevance', 'Recall', 'Precision', 'Consistency', 'Bias Detection']
selected_metrics = st.multiselect("Select Metrics to Analyze", metrics, default=metrics)
if selected_metrics:
# Set a threshold for "poor performance"
threshold = st.slider("Performance Threshold (%)", min_value=0, max_value=100, value=50)
# Filter data where any of the selected metrics are below the threshold
mask = df[selected_metrics].lt(threshold).any(axis=1)
worst_performing_df = df[mask]
if not worst_performing_df.empty:
st.write(f"Found {len(worst_performing_df)} evaluations below the threshold of {threshold}% in the selected metrics.")
# Display the worst-performing prompts and their metrics
st.write("### Worst Performing Evaluations")
display_columns = ['prompt', 'response'] + selected_metrics + ['timestamp']
worst_performing_display_df = worst_performing_df[display_columns].copy()
worst_performing_display_df['timestamp'] = worst_performing_display_df['timestamp'].dt.strftime('%Y-%m-%d %H:%M:%S')
# Apply styling to highlight low scores
def highlight_low_scores(val):
if isinstance(val, float):
if val < threshold:
return 'background-color: red; color: white'
return ''
styled_worst_df = worst_performing_display_df.style.applymap(highlight_low_scores, subset=selected_metrics)
styled_worst_df = styled_worst_df.format({metric: "{:.2f}%" for metric in selected_metrics})
st.dataframe(
styled_worst_df.set_properties(**{
'text-align': 'left',
'border': '1px solid #ddd'
}).set_table_styles([
{'selector': 'th', 'props': [('background-color', '#333'), ('color', 'white')]},
{'selector': 'td', 'props': [('vertical-align', 'top')]}
]),
use_container_width=True
)
# Analyze the worst-performing slices based on prompt characteristics
st.write("### Analysis by Prompt Length")
# Add a column for prompt length
worst_performing_df['Prompt Length'] = worst_performing_df['prompt'].apply(lambda x: len(x.split()))
# Define bins for prompt length ranges
bins = [0, 5, 10, 20, 50, 100, 1000]
labels = ['0-5', '6-10', '11-20', '21-50', '51-100', '100+']
worst_performing_df['Prompt Length Range'] = pd.cut(worst_performing_df['Prompt Length'], bins=bins, labels=labels, right=False)
# Group by 'Prompt Length Range' and calculate average metrics
group_metrics = worst_performing_df.groupby('Prompt Length Range')[selected_metrics].mean().reset_index()
# Display the average metrics per prompt length range
st.write("#### Average Metrics per Prompt Length Range")
group_metrics = group_metrics.sort_values('Prompt Length Range')
st.dataframe(group_metrics.style.format({metric: "{:.2f}%" for metric in selected_metrics}))
# Visualization of average metrics per prompt length range
st.write("#### Visualization of Metrics by Prompt Length Range")
melted_group_metrics = group_metrics.melt(id_vars='Prompt Length Range', value_vars=selected_metrics, var_name='Metric', value_name='Average Score')
fig = px.bar(
melted_group_metrics,
x='Prompt Length Range',
y='Average Score',
color='Metric',
barmode='group',
title='Average Metric Scores by Prompt Length Range',
labels={'Average Score': 'Average Score (%)'},
height=600
)
st.plotly_chart(fig, use_container_width=True)
# Further analysis: show counts of worst-performing evaluations per model
st.write("### Worst Performing Evaluations per Model")
model_counts = worst_performing_df['model_name'].value_counts().reset_index()
model_counts.columns = ['Model Name', 'Count of Worst Evaluations']
st.dataframe(model_counts)
# Allow user to download the worst-performing data
csv = worst_performing_df.to_csv(index=False)
st.download_button(
label="Download Worst Performing Data as CSV",
data=csv,
file_name='worst_performing_data.csv',
mime='text/csv',
)
else:
st.info("No evaluations found below the specified threshold.")
else:
st.warning("Please select at least one metric to analyze.")
else:
st.info("No evaluation results available for the selected model.")
except Exception as e:
st.error(f"Error processing data from database: {str(e)}")
st.error("Detailed error information:")
st.error(traceback.format_exc())
st.stop()
elif app_mode == "Model Upload":
st.title("Upload Your Model")
model_type = st.radio("Select Model Type", ["Custom", "Named"]) # {{ edit_6: Select model type }}
uploaded_file = st.file_uploader("Choose a model file", type=[".pt", ".h5", ".bin"]) if model_type == "custom" else None
if st.button("Upload Model"):
if model_type == "custom" and uploaded_file is not None:
result = upload_model(uploaded_file, st.session_state.user, model_type="custom")
st.success(result)
elif model_type == "named":
result = upload_model(None, st.session_state.user, model_type="named")
st.success(result)
else:
st.error("Please upload a valid model file for Custom models.")
elif app_mode == "Evaluation":
st.title("Evaluate Your Model")
st.write("### Select Model and Evaluation Metrics")
# Fetch the user from the database
user = users_collection.find_one({"username": st.session_state.user})
if user is None:
st.error("User not found in the database.")
st.stop()
user_models = user.get("models", [])
if not user_models:
st.error("You have no uploaded models. Please upload a model first.")
else:
# {{ edit_1: Display model_name instead of model_id }}
model_identifier = st.selectbox(
"Choose a Model to Evaluate",
[model['model_name'] if model['model_name'] else model['model_id'] for model in user_models]
)
# {{ edit_2: Remove metrics selection and set fixed metrics }}
fixed_metrics = ["Accuracy", "Hallucination", "Groundedness", "Relevance", "Recall", "Precision", "Consistency", "Bias Detection"]
st.write("### Evaluation Metrics")
st.write(", ".join(fixed_metrics))
# Modify the evaluation function call to use fixed_metrics
if st.button("Start Evaluation"):
with st.spinner("Evaluation in progress..."):
# {{ edit_3: Use fixed_metrics instead of user-selected metrics }}
results = evaluate_model(model_identifier, fixed_metrics, st.session_state.user)
# Fetch the current model document
current_model = next((m for m in user_models if (m['model_name'] == model_identifier) or (m['model_id'] == model_identifier)), None)
if current_model:
save_results(st.session_state.user, current_model, prompt, context, response, results) # {{ edit_21: Pass current_model to save_results }}
st.success("Evaluation Completed!")
st.json(results)
else:
st.error("Selected model not found.")
elif app_mode == "Prompt Testing":
st.title("Prompt Testing")
model_selection_option = st.radio("Select Model Option:", ["Choose Existing Model", "Add New Model"])
if model_selection_option == "Choose Existing Model":
user = users_collection.find_one({"username": st.session_state.user})
user_models = user.get("models", [])
if not user_models:
st.error("You have no uploaded models. Please upload a model first.")
else:
model_options = [
f"{model['model_name']} ({model.get('model_type', 'Unknown').capitalize()})"
for model in user_models
]
selected_model = st.selectbox("Select a Model for Testing", model_options)
model_name = selected_model.split(" (")[0]
model_type = selected_model.split(" (")[1].rstrip(")")
else:
# Code for adding a new model (unchanged)
...
st.subheader("Input for Model Testing")
# For simple models, we'll use a single JSON file
if model_type.lower() == "simple":
st.write("For simple models, please upload a single JSON file containing prompts, contexts, and responses.")
json_file = st.file_uploader("Upload Test Data JSON", type=["json"])
if json_file is not None:
try:
test_data = json.load(json_file)
st.success("Test data JSON file uploaded successfully!")
# Display a preview of the test data
st.write("Preview of test data:")
st.json(test_data[:3] if len(test_data) > 3 else test_data)
except json.JSONDecodeError:
st.error("Invalid JSON format. Please check your file.")
else:
test_data = None
else:
# For other model types, keep the existing separate inputs for context and questions
context_input_method = st.radio("Choose context input method:", ["Text Input", "File Upload"])
if context_input_method == "Text Input":
context_dataset = st.text_area("Enter Context Dataset (txt):", height=200)
else:
context_file = st.file_uploader("Upload Context Dataset", type=["txt"])
if context_file is not None:
context_dataset = context_file.getvalue().decode("utf-8")
st.success("Context file uploaded successfully!")
else:
context_dataset = None
questions_input_method = st.radio("Choose questions input method:", ["Text Input", "File Upload"])
if questions_input_method == "Text Input":
questions_json = st.text_area("Enter Questions (JSON format):", height=200)
else:
questions_file = st.file_uploader("Upload Questions JSON", type=["json"])
if questions_file is not None:
questions_json = questions_file.getvalue().decode("utf-8")
st.success("Questions file uploaded successfully!")
else:
questions_json = None
if st.button("Run Test"):
if not model_name:
st.error("Please select or add a valid Model.")
elif model_type.lower() == "simple" and test_data is None:
st.error("Please upload a valid test data JSON file.")
elif model_type.lower() != "simple" and (not context_dataset or not questions_json):
st.error("Please provide both context dataset and questions JSON.")
else:
try:
selected_model = next(
(m for m in user_models if m['model_name'] == model_name),
None
)
if selected_model:
with st.spinner("Starting evaluations..."):
if model_type.lower() == "simple":
evaluation_thread = threading.Thread(
target=run_custom_evaluations,
args=(test_data, selected_model, st.session_state.user)
)
else:
questions = json.loads(questions_json)
evaluation_thread = threading.Thread(
target=run_custom_evaluations,
args=((context_dataset, questions), selected_model, st.session_state.user)
)
evaluation_thread.start()
st.success("Evaluations are running in the background. You can navigate away or close the site.")
else:
st.error("Selected model not found.")
except json.JSONDecodeError:
st.error("Invalid JSON format. Please check your input.")
elif app_mode == "Manage Models":
st.title("Manage Your Models")
# Fetch the user from the database
user = users_collection.find_one({"username": st.session_state.user})
if user is None:
st.error("User not found in the database.")
st.stop()
user_models = user.get("models", [])
# Update existing models to ensure they have a model_type
for model in user_models:
if 'model_type' not in model:
model['model_type'] = 'simple' # Default to 'simple' for existing models
users_collection.update_one(
{"username": st.session_state.user},
{"$set": {"models": user_models}}
)
st.subheader("Add a New Model")
model_type = st.radio("Select Model Type:", ["Simple Model", "Custom Model"])
if model_type == "Simple Model":
new_model_name = st.text_input("Enter New Model Name:")
if st.button("Add Simple Model") or st.button("Add Custom Model"):
if new_model_name or selected_custom_model:
model_id = f"{st.session_state.user}_model_{int(datetime.now().timestamp())}"
model_data = {
"model_id": model_id,
"model_name": new_model_name if model_type == "Simple Model" else selected_custom_model,
"model_type": "simple" if model_type == "Simple Model" else "custom",
"file_path": None,
"model_link": None,
"uploaded_at": datetime.now(),
"context": None # We'll update this when running evaluations
}
users_collection.update_one(
{"username": st.session_state.user},
{"$push": {"models": model_data}}
)
st.success(f"Model '{model_data['model_name']}' added successfully as {model_id}!")
else:
st.error("Please enter a valid model name or select a custom model.")
else: # Custom Model
custom_model_options = ["gpt-4o", "gpt-4o-mini"]
selected_custom_model = st.selectbox("Select Custom Model:", custom_model_options)
if st.button("Add Custom Model"):
model_id = f"{st.session_state.user}_model_{int(datetime.now().timestamp())}"
users_collection.update_one(
{"username": st.session_state.user},
{"$push": {"models": {
"model_id": model_id,
"model_name": selected_custom_model,
"model_type": "custom",
"file_path": None,
"model_link": None,
"uploaded_at": datetime.now()
}}}
)
st.success(f"Custom Model '{selected_custom_model}' added successfully as {model_id}!")
st.markdown("---")
if user_models:
st.subheader("Your Models")
for model in user_models:
st.markdown(f"**Model ID:** {model['model_id']}")
st.write(f"**Model Type:** {model.get('model_type', 'simple').capitalize()}")
if model.get("model_name"):
st.write(f"**Model Name:** {model['model_name']}")
if model.get("file_path"):
st.write(f"**File Path:** {model['file_path']}")
st.write(f"**Uploaded at:** {model['uploaded_at']}")
# Add delete option
if st.button(f"Delete {model['model_id']}"):
# Delete the model file if exists and it's a Custom model
if model['file_path'] and os.path.exists(model['file_path']):
os.remove(model['file_path'])
# Remove model from user's models list
users_collection.update_one(
{"username": st.session_state.user},
{"$pull": {"models": {"model_id": model['model_id']}}}
)
st.success(f"Model {model['model_id']} deleted successfully!")
else:
st.info("You have no uploaded models.")
elif app_mode == "History": # {{ edit_add: Enhanced History UI }}
st.title("History")
st.write("### Your Evaluation History")
try:
# Fetch all evaluation results for the current user from MongoDB
user_results = list(results_collection.find({"username": st.session_state.user}).sort("timestamp", -1))
if user_results:
# Convert results to a pandas DataFrame
df = pd.DataFrame(user_results)
# Extract prompt text using the helper function
df['prompt'] = df['prompt'].apply(extract_prompt_text)
# Normalize the evaluation JSON into separate columns
eval_df = df['evaluation'].apply(pd.Series)
for metric in ["Accuracy", "Hallucination", "Groundedness", "Relevance", "Recall", "Precision", "Consistency", "Bias Detection"]:
if metric in eval_df.columns:
df[metric + " Score"] = eval_df[metric].apply(lambda x: x.get('score', 0) * 100 if isinstance(x, dict) else 0)
df[metric + " Explanation"] = eval_df[metric].apply(lambda x: x.get('explanation', '') if isinstance(x, dict) else '')
else:
df[metric + " Score"] = 0
df[metric + " Explanation"] = ""
# Select relevant columns to display
display_df = df[[
"timestamp", "model_name", "prompt", "context", "response",
"Accuracy Score", "Hallucination Score", "Groundedness Score",
"Relevance Score", "Recall Score", "Precision Score",
"Consistency Score", "Bias Detection Score"
]]
# Rename columns for better readability
display_df = display_df.rename(columns={
"timestamp": "Timestamp",
"model_name": "Model Name",
"prompt": "Prompt",
"context": "Context",
"response": "Response",
"Accuracy Score": "Accuracy (%)",
"Hallucination Score": "Hallucination (%)",
"Groundedness Score": "Groundedness (%)",
"Relevance Score": "Relevance (%)",
"Recall Score": "Recall (%)",
"Precision Score": "Precision (%)",
"Consistency Score": "Consistency (%)",
"Bias Detection Score": "Bias Detection (%)"
})
# Convert timestamp to a readable format
display_df['Timestamp'] = pd.to_datetime(display_df['Timestamp']).dt.strftime('%Y-%m-%d %H:%M:%S')
st.subheader("Evaluation Results")
# Display the DataFrame with enhanced styling
st.dataframe(
display_df.style.set_properties(**{
'background-color': '#f0f8ff',
'color': '#333',
'border': '1px solid #ddd'
}).set_table_styles([
{'selector': 'th', 'props': [('background-color', '#f5f5f5'), ('text-align', 'center')]},
{'selector': 'td', 'props': [('text-align', 'left'), ('vertical-align', 'top')]}
]).format({
"Accuracy (%)": "{:.2f}",
"Hallucination (%)": "{:.2f}",
"Groundedness (%)": "{:.2f}",
"Relevance (%)": "{:.2f}",
"Recall (%)": "{:.2f}",
"Precision (%)": "{:.2f}",
"Consistency (%)": "{:.2f}",
"Bias Detection (%)": "{:.2f}"
}), use_container_width=True
)
else:
st.info("You have no evaluation history yet.")
except Exception as e:
st.error(f"Error fetching history data: {e}")
# Add a footer
st.sidebar.markdown("---")
st.sidebar.info("LLM Evaluation System - v0.2")