Traces

​

Overview

​

What are Traces?

​

Key Features

​

Distributed Tracing

• Trace ID propagation
• Parent-child relationships
• Cross-Project tracing
• Timeline visualization

​

Span Details

• Function name and duration
• Input and output data
• Error messages and stack traces
• Metadata and tags
• Resource consumption

​

LLM-Specific Tracing

• Prompt templates and variables
• System and user messages
• Tool calls considered
• Tool execution results
• Model responses
• Token usage breakdown
• Reasoning/chain-of-thought

​

Performance Analysis

• Span duration breakdown
• Critical path analysis
• Concurrency visualization
• Resource usage hotspots

​

Cost Attribution

• Cost per span
• LLM token costs
• API call costs
• Aggregate by component

​

Example Trace

Trace ID: trace_abc123xyz
Duration: 3.2s
Cost: $0.04

└─ Flow: customer_support_flow (3.2s)
   ├─ Action: validate_input (0.1s, $0)
   │  └─ Input: {"message": "I need help"}
   │  └─ Output: {"valid": true}
   │
   ├─ Agent: support_agent (2.8s, $0.038)
   │  ├─ Prompt construction (0.01s)
   │  │  └─ Template: support_agent_system.txt
   │  │  └─ Variables: {company_name, policies}
   │  │
   │  ├─ LLM call: gpt-4 (1.5s, $0.030)
   │  │  └─ Prompt tokens: 850
   │  │  └─ Completion tokens: 200
   │  │  └─ Response: "I understand you need help..."
   │  │  └─ Tool calls: [check_account_status]
   │  │
   │  ├─ Tool: check_account_status (1.2s, $0)
   │  │  └─ Action: fetch_user_data
   │  │  └─ Duration: 1.2s
   │  │  └─ Result: {status: "active"}
   │  │
   │  └─ LLM call: gpt-4 (0.09s, $0.008)
   │     └─ With tool results
   │     └─ Final response generated
   │
   └─ Action: format_output (0.3s, $0)
      └─ Markdown formatting applied

​

Visualization

​

Timeline View

• Horizontal bars showing duration
• Color-coded by type
• Nested structure
• Click to expand/collapse

​

Graph View

• Components as nodes
• Data flow as edges
• Highlight critical path
• Filter by type

​

Flamegraph

• Hierarchical time distribution
• Identify hot spots
• Interactive drill-down

​

Search and Filter

status:error AND duration:>5s AND agent:customer_support

• Status (success, error, timeout)
• Duration range
• Cost range
• Component type
• Tag or metadata
• Time range

​

Trace Sampling

• All traces: Full observability (dev/staging)
• Sampled: Random sample (production, high volume)
• Error-only: Trace only failures
• Smart sampling: Trace slow or expensive executions

​

Integrations

​

OpenTelemetry

• Export to Jaeger
• Export to Zipkin
• Export to Datadog
• Export to New Relic

​

Custom Backends

• S3/Cloud storage
• Your own database
• Analytics platforms
• SIEM systems

​

Real-Time Streaming

• Watch executions in real-time
• Stream to dashboard
• WebSocket updates
• Tail logs live

​

Trace Context Propagation

# Your app
trace_id = generate_trace_id()

# Call Triform
response = triform.execute(
    project="my-project",
    payload={"data": "..."},
    trace_id=trace_id  # Propagate context
)

# Traces connected across systems

​

Retention

• Free: 7 days
• Pro: 30 days
• Enterprise: 90 days or custom

​

Use Cases

​

Debugging

• Why did execution fail?
• Where is the slowdown?
• What input caused this?

​

Optimization

• Identify slowest steps
• Find redundant calls
• Optimize critical path

​

Cost Management

• Which components are expensive?
• Can we use cheaper models?
• Where to add caching?

​

Compliance

• What data was accessed?
• Which models were used?
• Who triggered execution?

​

API

from triform import Traces

# Get trace
trace = Traces.get("trace_abc123")

# List traces
traces = Traces.list(
    project="my-project",
    start_time="2025-10-01",
    end_time="2025-10-02",
    status="error"
)

# Export traces
Traces.export(
    trace_ids=["trace_1", "trace_2"],
    format="json",
    destination="s3://my-bucket/traces/"
)

​

Pricing

​

Timeline

​

Get Notified

​

Questions?

• Join Discord #observability channel
• Email: product@triform.ai