Observations

Observations in the OuterBox AI Observability platform provide detailed insights into the individual components and steps that make up each interaction with your P3 AI Web Assistant. They offer a granular view of how your assistant processes and responds to user inquiries.

What Are Observations?

Observations represent the specific processing steps and components that occur during an interaction with your AI assistant. They include:

• Individual processing steps within a conversation
• Model calls and their specific parameters
• Response generation details
• Timing and performance metrics for each step

Think of observations as the building blocks that make up each trace, allowing you to see exactly how your assistant processes and responds to queries.

Types of Observations

The OuterBox AI Observability platform tracks several types of observations:

Events

Simple, discrete occurrences during processing, such as:

• When a user submits a question
• When specific processing steps begin or end
• When certain conditions or thresholds are met

Spans

Time periods representing specific processing activities, such as:

• The duration of a particular analysis step
• The time spent retrieving information from your knowledge base
• The overall response generation time

Generations

Special spans that represent the creation of AI-generated content, including:

• The actual text generation process
• Model parameters used for the generation
• Token usage and associated costs

How to Use Observations Effectively

Accessing Observations

1. Navigate to the "Observations" section in the left sidebar
2. Browse the list of recent observations across all interactions
3. Filter observations by type, duration, or other attributes

Analyzing Observations

When examining observations, you can:

• See exactly which models and parameters were used
• Identify which processing steps took the most time
• Understand how your assistant retrieved and processed information
• Track resource usage and costs at a granular level

Common Use Cases for Observations

• Performance Optimization: Identify processing bottlenecks or inefficiencies
• Cost Analysis: See exactly which components are driving your usage costs
• Technical Troubleshooting: Debug specific issues in the assistant's processing pipeline
• Model Evaluation: Compare how different models or parameters perform for similar queries

Best Practices for Working with Observations

• Focus on Patterns: Look for consistent patterns across multiple observations
• Compare Similar Queries: See how your assistant handles similar questions differently
• Track Changes Over Time: Monitor how observations change after updates to your assistant
• Connect to User Experience: Relate technical observations to the actual user experience

Filtering and Analyzing Observations

The platform provides powerful tools for working with observations:

• Type Filters: Focus on specific types of observations (events, spans, generations)
• Duration Filters: Identify particularly fast or slow processing steps
• Model Filters: Compare observations across different AI models
• Cost Filters: Focus on the most resource-intensive operations

Using Observations for Optimization

To get the most value from observations:

1. Identify the most common or critical user queries
2. Analyze the observations for these queries to understand processing patterns
3. Look for opportunities to optimize response generation or information retrieval
4. Implement changes and monitor the impact through subsequent observations

Observations provide the most detailed technical view of your assistant's operation, making them invaluable for optimization and troubleshooting.