Root Cause Analysis

Identify the root causes of performance changes and anomalies

Overview

When drift or anomalies are detected, TuringPulse helps identify the root cause by correlating changes across deployments, configurations, and external factors.

Attribution Sources

TuringPulse tracks and correlates multiple change sources:

Deployments - Code and model version changes
Prompts - System prompt modifications
Configurations - Parameter changes (temperature, max_tokens)
External APIs - Third-party service changes
Data Distribution - Input pattern changes

Change Timeline

The Change Timeline view shows all tracked changes overlaid with your metrics:

Deployment markers with commit information
Configuration change events
Anomaly and drift events
Metric trend lines

Tracking Deployments

from turingpulse_sdk import register_deploy

# Track deployment at startup
register_deploy(
    workflow_id="my-agent",
    version="1.2.3",
    git_sha="abc123",
    commit_message="Updated system prompt, increased max_tokens",
    auto_detect=True,
)

Tracking Prompt Changes

from turingpulse_sdk import register_deploy

# Track prompt modification as a deploy event
register_deploy(
    workflow_id="my-agent",
    version="prompt-v2",
    commit_message="Added safety guidelines to system prompt",
    metadata={
        "prompt_id": "system_prompt_v1",
        "change_type": "prompt_update",
    },
)

💡

Automatic Detection

When using LangChain or LangGraph prompts, changes are automatically detected and tracked without manual instrumentation.

Attribution Rules

Define custom attribution rules in the UI to automatically correlate specific types of changes with observed impacts:

Model version changes → latency/cost impacts
Prompt changes → quality metric impacts
Temperature changes → output variance

Best Practices

Track all deployments with semantic version numbers
Include commit SHA for traceability
Document changes in deployment metadata
Review attribution insights after each deployment
Set up alerts for high-impact changes