Span Class

The Span class represents a specific operation within a trace. Spans are the building blocks of observability, tracking individual steps like LLM calls, tool executions, database queries, or any custom operation in your LLM pipeline.

Overview

A span captures:

Operation details: name, timing, status
Content: input/output data for different span types (Model, Tool, Retrieval, etc.)
Metadata: tags, attributes for filtering and analysis
Hierarchy: parent-child relationships for nested operations
Evaluation: optional evaluator execution

Common span types:

Model - LLM inference calls
ModelStream - Streaming LLM responses
Tool - Function/API executions
Retrieval - RAG and vector searches
Embeddings - Embedding generation
Function - Custom application logic
Guardrail - Safety/compliance checks
Other - Any custom operation

Creation

Create spans from a trace or parent span:

// From trace
const span = trace.logSpan({
  name: 'LLM Call',
  tags: ['llm']
});

// From parent span (nested)
const childSpan = parentSpan.logSpan({
  name: 'Tool Execution',
  tags: ['tool']
});

Properties

span

span: CreateLogSpanRequest

The underlying span request object containing all span data. Structure:

{
  projectId: string;
  span: {
    name: string;
    status: SpanStatus;
    referenceId: string;
    traceId?: string;
    traceReferenceId: string;
    parentReferenceId?: string;
    sessionId?: string;
    promptId?: string;
    deploymentId?: string;
    runEvaluation?: boolean;
    tags?: string[];
    attributes?: Record<string, string | number | boolean>;
    content: LogSpanContent;
    startedAt: number;
    endedAt: number;
  };
}

Methods

logSpan()

Create a nested child span under this span.

logSpan(options: LogSpanOptions): Span

This allows you to create hierarchical span relationships like:

Trace
  └─ Parent Span
      ├─ Child Span 1
      └─ Child Span 2
          └─ Grandchild Span

Parameters

options

object

required

Same parameters as Trace.logSpan(). See Trace.logSpan() for details.

Returns

span

Span

A new child Span instance linked to this parent span.

Example

const parentSpan = trace.logSpan({ name: 'RAG Pipeline' });

// Create child spans
const embeddingSpan = parentSpan.logSpan({
  name: 'Generate Embedding',
  tags: ['embedding']
});
embeddingSpan.end();

const retrievalSpan = parentSpan.logSpan({
  name: 'Vector Search',
  tags: ['retrieval']
});
retrievalSpan.end();

parentSpan.end();

update()

Update span metadata and content.

update(updates: SpanUpdates): this

Parameters

updates

object

required

Show properties

name

string

Update the span name.

status

SpanStatus

Update SpanStatus: 'success' | 'failure' | 'aborted' | 'cancelled' | 'unknown'

content

LogSpanContent

Update the span LogSpanContent (input/output data).

runEvaluation

boolean

Update whether to run evaluators.

Returns

Returns this for method chaining.

Examples

const span = trace.logSpan({ name: 'API Call' });

try {
  await callAPI();
  span.update({ status: 'success' });
} catch (error) {
  span.update({ status: 'failure' });
}

span.end();

const span = trace.logSpan({ name: 'LLM Call' });

const gatewayResponse = await gateway.completeChat({ model, messages, tools });

span.update({
  status: 'success',
  content: {
    type: 'Model',
    provider: 'openai',
    model: 'gpt-4o',
    input: JSON.stringify(gatewayResponse.provider.request),
    output: JSON.stringify(gatewayResponse.provider.response),
    cost: 0.002
  }
});

span.end();

trace.logSpan({ name: 'Quick Op' })
  .update({ status: 'unknown' })
  .update({ tags: ['fast'] })
  .update({ status: 'success' })
  .end();

end()

Mark the span as complete and ready to be flushed.

end(): string | undefined

Behavior

Sets endedAt timestamp
Marks the span as ready in the monitor’s buffer
Recursively ends all child spans
Returns the span’s reference ID

Always call end() on your spans! Spans that are never ended will never be flushed to the API.

Returns

referenceId

string | undefined

The span’s reference ID for correlation.

Example

const span = trace.logSpan({ name: 'Operation' });

try {
  await doWork();
  span.update({ status: 'success' });
} finally {
  span.end(); // Always called
}

Span Content Types

Different content types for different operations. See LogSpanContent Type for complete documentation.

Model (LLM Inference)

For standard LLM completions (LogSpanModelContent).

span.update({
  content: {
    type: 'Model',
    provider: 'openai',           // Provider name
    model: 'gpt-4o',              // Model identifier
    variables: {                   // Optional prompt variables
      userName: 'John',
      context: 'support'
    },
    cost: 0.002,                  // Optional cost in USD
    input: JSON.stringify([       // Input messages
      { role: 'user', content: 'Hello' }
    ]),
    output: JSON.stringify({      // Output message
      role: 'assistant',
      content: 'Hi there!'
    }),
    expected: JSON.stringify({    // Optional expected output for evaluation
      role: 'assistant',
      content: 'Hello! How can I help?'
    })
  }
});

The variables field is typed as LogSpanVariable | null in the SDK. LogSpanVariable is a record type, so you can pass a plain object as shown above.

ModelStream (Streaming LLM)

For streaming LLM responses.

let chunks = '';
for await (const chunk of stream) {
  chunks += chunk;
}

span.update({
  content: {
    type: 'ModelStream',
    provider: 'anthropic',
    model: 'claude-3-opus',
    input: JSON.stringify(messages),
    output: chunks,                        // Raw stream chunks
    aggregateOutput: JSON.stringify({      // Final aggregated response
      role: 'assistant',
      content: fullResponse
    }),
    cost: 0.005
  }
});

Tool (Function/API Call)

For tool executions and API calls.

span.update({
  content: {
    type: 'Tool',
    input: JSON.stringify({
      function: 'getWeather',
      city: 'San Francisco'
    }),
    output: JSON.stringify({
      temperature: 72,
      conditions: 'sunny',
      humidity: 65
    })
  }
});

Retrieval (RAG/Vector Search)

For document retrieval and vector searches.

span.update({
  content: {
    type: 'Retrieval',
    input: JSON.stringify({
      query: 'What is machine learning?',
      topK: 5,
      filters: { category: 'AI' }
    }),
    output: JSON.stringify({
      documents: [
        { id: 'doc1', score: 0.95, text: '...' },
        { id: 'doc2', score: 0.89, text: '...' }
      ]
    })
  }
});

Embeddings

For embedding generation.

span.update({
  content: {
    type: 'Embeddings',
    input: JSON.stringify({
      texts: ['text1', 'text2', 'text3'],
      model: 'text-embedding-3-large'
    }),
    output: JSON.stringify({
      embeddings: [[0.1, 0.2, ...], [0.3, 0.4, ...]],
      dimensions: 3072
    })
  }
});

Function (Custom Logic)

For custom application logic.

span.update({
  content: {
    type: 'Function',
    input: JSON.stringify({
      operation: 'processData',
      params: { id: 123 }
    }),
    output: JSON.stringify({
      result: 'success',
      itemsProcessed: 42
    })
  }
});

Guardrail (Safety Check)

For safety and compliance checks.

span.update({
  content: {
    type: 'Guardrail',
    input: JSON.stringify({
      text: 'User input to check...',
      checks: ['toxicity', 'pii', 'harm']
    }),
    output: JSON.stringify({
      safe: true,
      scores: {
        toxicity: 0.05,
        pii: 0.01,
        harm: 0.02
      }
    })
  }
});

Other (Custom)

For any other operation type.

span.update({
  content: {
    type: 'Other',
    input: JSON.stringify({ custom: 'input' }),
    output: JSON.stringify({ custom: 'output' })
  }
});

Complete Examples

Example 1: LLM Call with Deployment

import { Adaline } from '@adaline/client';
import { Gateway } from '@adaline/gateway';
import { OpenAI } from '@adaline/openai';

const adaline = new Adaline();
const gateway = new Gateway();
const openaiProvider = new OpenAI();
const monitor = adaline.initMonitor({ projectId: 'my-project' });

async function generateResponse(userId: string, message: string) {
  // Get deployment
  const deployment = await adaline.getLatestDeployment({
    promptId: 'chat-prompt',
    deploymentEnvironmentId: 'environment_abc123'
  });

  const trace = monitor.logTrace({
    name: 'Chat Completion',
    sessionId: userId
  });

  const span = trace.logSpan({
    name: 'LLM Completion',
    promptId: deployment.promptId,
    deploymentId: deployment.id,
    runEvaluation: true,
    tags: ['llm', 'openai']
  });

  try {
    const model = openaiProvider.chatModel({
      modelName: deployment.prompt.config.model,
      apiKey: process.env.OPENAI_API_KEY!
    });

    const gatewayResponse = await gateway.completeChat({
      model,
      config: deployment.prompt.config.settings,
      messages: [
        ...deployment.prompt.messages,
        { role: 'user', content: [{ modality: 'text', value: message }] }
      ],
      tools: deployment.prompt.tools
    });

    const reply = String(gatewayResponse.response.messages[0].content[0].value);

    span.update({
      status: 'success',
      content: {
        type: 'Model',
        provider: deployment.prompt.config.providerName,
        model: deployment.prompt.config.model,
        input: JSON.stringify(gatewayResponse.provider.request),
        output: JSON.stringify(gatewayResponse.provider.response)
      },
    });

    return reply;

  } catch (error) {
    span.update({
      status: 'failure',
      attributes: {
        error: error instanceof Error ? error.message : String(error)
      }
    });
    throw error;

  } finally {
    span.end();
    trace.end();
  }
}

Example 2: Nested RAG Pipeline

import { ChromaClient } from 'chromadb';

const chroma = new ChromaClient();

async function ragPipeline(question: string) {
  const trace = monitor.logTrace({
    name: 'RAG Pipeline',
    tags: ['rag']
  });

  const pipelineSpan = trace.logSpan({
    name: 'RAG Workflow',
    tags: ['pipeline']
  });

  try {
    // Step 1: Embedding
    const embedSpan = pipelineSpan.logSpan({
      name: 'Generate Embedding',
      tags: ['embedding']
    });

    const embedResponse = await openai.embeddings.create({
      model: 'text-embedding-3-large',
      input: question
    });
    const embedding = embedResponse.data[0].embedding;

    embedSpan.update({
      status: 'success',
      content: {
        type: 'Embeddings',
        input: JSON.stringify({ text: question }),
        output: JSON.stringify({
          dimensions: embedding.length,
          model: 'text-embedding-3-large'
        })
      }
    });
    embedSpan.end();

    // Step 2: Retrieval
    const retrievalSpan = pipelineSpan.logSpan({
      name: 'Vector Search',
      tags: ['retrieval', 'chromadb']
    });

    const collection = await chroma.getCollection({ name: 'docs' });
    const results = await collection.query({
      queryEmbeddings: [embedding],
      nResults: 5
    });

    retrievalSpan.update({
      status: 'success',
      content: {
        type: 'Retrieval',
        input: JSON.stringify({
          query: question,
          collection: 'docs',
          topK: 5
        }),
        output: JSON.stringify({
          documentIds: results.ids[0],
          distances: results.distances?.[0],
          documents: results.documents[0]
        })
      },
      attributes: {
        documentsFound: results.ids[0].length
      }
    });
    retrievalSpan.end();

    // Step 3: LLM Generation
    const llmSpan = pipelineSpan.logSpan({
      name: 'Generate Answer',
      runEvaluation: true,
      tags: ['llm', 'answer']
    });

    const context = results.documents[0].join('\n\n');

    const llmModel = openaiProvider.chatModel({
      modelName: 'gpt-4o',
      apiKey: process.env.OPENAI_API_KEY!
    });

    const gatewayResponse = await gateway.completeChat({
      model: llmModel,
      messages: [
        { role: 'system', content: [{ modality: 'text', value: 'Answer based on the provided context.' }] },
        { role: 'user', content: [{ modality: 'text', value: `Context:\n${context}\n\nQuestion: ${question}` }] }
      ]
    });

    const answer = String(gatewayResponse.response.messages[0].content[0].value);

    llmSpan.update({
      status: 'success',
      content: {
        type: 'Model',
        provider: 'openai',
        model: 'gpt-4o',
        input: JSON.stringify(gatewayResponse.provider.request),
        output: JSON.stringify(gatewayResponse.provider.response)
      },
      attributes: {
        contextLength: context.length
      }
    });
    llmSpan.end();

    pipelineSpan.update({ status: 'success' });
    pipelineSpan.end();

    trace.update({ status: 'success' });

    return answer;

  } catch (error) {
    pipelineSpan.update({ status: 'failure' });
    trace.update({ status: 'failure' });
    throw error;

  } finally {
    trace.end();
  }
}

Example 3: Guardrail Check

async function checkContent(text: string) {
  const trace = monitor.logTrace({ name: 'Content Check' });

  const guardrailSpan = trace.logSpan({
    name: 'Safety Guardrail',
    tags: ['guardrail', 'safety']
  });

  try {
    // Call safety API
    const response = await fetch('https://api.safety.com/v1/check', {
      method: 'POST',
      body: JSON.stringify({ text })
    });
    const result = await response.json();

    const isSafe = result.toxicity < 0.1 && result.harm < 0.1;

    guardrailSpan.update({
      status: 'success',
      content: {
        type: 'Guardrail',
        input: JSON.stringify({ text }),
        output: JSON.stringify({
          safe: isSafe,
          scores: {
            toxicity: result.toxicity,
            harm: result.harm,
            pii: result.pii
          }
        })
      },
      attributes: {
        textLength: text.length,
        safe: isSafe
      }
    });

    guardrailSpan.end();
    trace.end();

    return isSafe;

  } catch (error) {
    guardrailSpan.update({ status: 'failure' });
    guardrailSpan.end();
    trace.end();
    throw error;
  }
}

Type Definitions

type SpanStatus = 
  | 'success'
  | 'failure'
  | 'aborted'
  | 'cancelled'
  | 'unknown';

type LogAttributesValue = string | number | boolean;

interface LogSpanOptions {
  name: string;
  status?: SpanStatus;
  referenceId?: string;
  promptId?: string;
  deploymentId?: string | null;
  runEvaluation?: boolean;
  tags?: string[];
  attributes?: Record<string, LogAttributesValue>;
  content?: LogSpanContent;
}

interface SpanUpdates {
  name?: string;
  status?: SpanStatus;
  content?: LogSpanContent;
  runEvaluation?: boolean;
  tags?: string[];
  attributes?: Record<string, LogAttributesValue>;
}

type LogSpanContent =
  | LogSpanModelContent
  | LogSpanModelStreamContent
  | LogSpanEmbeddingsContent
  | LogSpanFunctionContent
  | LogSpanToolContent
  | LogSpanGuardrailContent
  | LogSpanRetrievalContent
  | LogSpanOtherContent;

Best Practices

1. Use Appropriate Content Types

// ✅ Good: Correct content type for operation
const llmSpan = trace.logSpan({ name: 'LLM' });
llmSpan.update({
  content: {
    type: 'Model',
    provider: 'openai',
    model: 'gpt-4o',
    input: '...',
    output: '...'
  }
});

const toolSpan = trace.logSpan({ name: 'API' });
toolSpan.update({
  content: {
    type: 'Tool',
    input: '...',
    output: '...'
  }
});

2. Always End Spans

// ✅ Good: Always end
const span = trace.logSpan({ name: 'Op' });
try {
  await work();
} finally {
  span.end(); // Always called
}

3. Use Attributes and Tags for Filtering

// ✅ Good: Descriptive tags
const span = trace.logSpan({
  name: 'OpenAI Call',
  tags: [
    'llm',
    'openai',
    'gpt-4o',
    'production',
    'premium-tier'
  ],
  attributes: {
    userId: 'user-123',
    region: 'us-east-1',
    cached: false,
  }
});

​Span Class

​Overview

​Creation

​Properties

​span

​Methods

​logSpan()

​Parameters

​Returns

​Example

​update()

​Parameters

​Returns

​Examples

​end()

​Behavior

​Returns

​Example

​Span Content Types

​Model (LLM Inference)

​ModelStream (Streaming LLM)

​Tool (Function/API Call)

​Retrieval (RAG/Vector Search)

​Embeddings

​Function (Custom Logic)

​Guardrail (Safety Check)

​Other (Custom)

​Complete Examples

​Example 1: LLM Call with Deployment

​Example 2: Nested RAG Pipeline

​Example 3: Guardrail Check

​Type Definitions

​Best Practices

​1. Use Appropriate Content Types

​2. Always End Spans

​3. Use Attributes and Tags for Filtering

Span Class

Overview

Creation

Properties

span

Methods

logSpan()

Parameters

Returns

Example

update()

Parameters

Returns

Examples

end()

Behavior

Returns

Example

Span Content Types

Model (LLM Inference)

ModelStream (Streaming LLM)

Tool (Function/API Call)

Retrieval (RAG/Vector Search)

Embeddings

Function (Custom Logic)

Guardrail (Safety Check)

Other (Custom)

Complete Examples

Example 1: LLM Call with Deployment

Example 2: Nested RAG Pipeline

Example 3: Guardrail Check

Type Definitions

Best Practices

1. Use Appropriate Content Types

2. Always End Spans

3. Use Attributes and Tags for Filtering