How Tags Work

Add tags during execution

Use the ABV SDK to attach one or more string tags to a trace. Tags can be added when creating a trace or updated later during execution.

abv.update_current_trace(tags=["rag", "beta", "gpt-4o"])

Tags appear in the dashboard

All tags attached to a trace are visible in the ABV UI. You’ll see them as clickable labels on each trace, making it easy to identify categories at a glance.

Filter traces by tags

Click any tag in the UI to filter your trace list. The filter shows only traces that include that specific tag, reducing noise and focusing your analysis.

Combine tags for precise filtering

Use multiple tag filters simultaneously to narrow down exactly what you need. For example, filter by both production and error to see only production errors.

Use tags in analytics and exports

Tags are included in all exports and available for grouping in custom dashboards. Use them to segment performance metrics, cost analysis, or error rates by any dimension you choose.

Why Use Tags?

Feature Flagging & Experiments

A/B testing a new prompt? Tag each variant to measure quality, latency, and cost separately.

variant = get_experiment_variant(user_id)
abv.update_current_trace(tags=[f"prompt:{variant}"])

Filter by tag to compare metrics side-by-side and make data-driven rollout decisions.Combine with metadata for richer analysis: use tags for simple categories (prompt:v1) and metadata for detailed attributes.

Version Tracking Across Deployments

Tag traces with your application version to isolate errors by deployment.

import os
APP_VERSION = os.getenv("APP_VERSION", "unknown")

@observe()
def handle_request(data):
    abv.update_current_trace(tags=[f"version:{APP_VERSION}"])
    # ... your logic ...

Filter by version:2.3.0 to see only new deployment errors and compare error rates across versions.Set version tags automatically via environment variables for complete deployment visibility.

Environment Separation

Separate dev, staging, and production traffic for clearer debugging.

ENVIRONMENT = os.getenv("ENVIRONMENT", "dev")

@observe()
def process_data(input_data):
    abv.update_current_trace(tags=[ENVIRONMENT, "data-pipeline"])
    # ... processing ...

Filter by production for real user traffic or staging for pre-release validation.For formal separation with access controls, see Environments.

Technique Identification (RAG, Few-Shot, etc.)

Tag traces by LLM technique (RAG, few-shot, chain-of-thought) to analyze performance and cost.

@observe()
def answer_question(question, context_docs):
    abv.update_current_trace(tags=["rag", "documentation"])
    # ... RAG implementation ...

@observe()
def classify_intent(user_input):
    abv.update_current_trace(tags=["few-shot", "classification"])
    # ... few-shot classification ...

Compare metrics to discover that RAG costs 3x more than few-shot or that chain-of-thought has higher latency but better accuracy.Optimize technique selection based on cost and performance data.

Error Categorization

Categorize errors for effective triage: rate limits, validation failures, or unexpected errors.

@observe()
def call_llm(prompt):
    try:
        response = llm.complete(prompt)
        return response
    except RateLimitError:
        abv.update_current_trace(tags=["error", "rate-limit"])
        raise
    except ValidationError:
        abv.update_current_trace(tags=["error", "validation"])
        raise
    except Exception:
        abv.update_current_trace(tags=["error", "unknown"])
        raise

Filter by error type to identify quota issues, bad input, or unexpected failures.Set up alerts on specific error tags to get notified only for critical issues.

User Cohorts & A/B Testing

Tag traces with user cohorts to measure adoption and performance across customer segments.

@observe()
def process_request(user_id):
    user = get_user(user_id)
    cohort_tags = [f"tier:{user.tier}", f"region:{user.region}"]
    abv.update_current_trace(tags=cohort_tags)
    # ... processing ...

Filter by tier:premium for paying customers or compare latency across regions. Segment cost analysis by customer tier.Avoid PII in tags—use cohort identifiers (tier:premium) not personal info (user:john@example.com).

Implementation Guide

Python SDK

With @observe() Decorator
With Manual Spans
Update Current Trace

The simplest approach for functions already decorated with @observe():

from abvdev import observe, ABV

# Initialize ABV client
abv = ABV(
    api_key="sk-abv-...",
    host="https://app.abv.dev"  # or "https://eu.app.abv.dev" for EU
)

@observe()
def process_document(doc_id, use_rag=True):
    # Add tags to categorize this trace
    tags = ["document-processing"]
    if use_rag:
        tags.append("rag")

    abv.update_current_trace(tags=tags)

    # ... your processing logic ...
    return result

# Call the function - tags are automatically attached
process_document("doc-123", use_rag=True)

When to use: For functions already using the @observe() decorator. Minimal code changes required.

When creating spans directly using context managers:

from abvdev import ABV

abv = ABV(
    api_key="sk-abv-...",
    host="https://app.abv.dev"
)

# Add tags when creating the root span
with abv.start_as_current_span(name="data-pipeline") as root_span:
    # Add tags to the trace
    root_span.update_trace(tags=["etl", "production", "v2.0"])

    # You can add more tags later from any span in the same trace
    with root_span.start_as_current_generation(
        name="llm-summary",
        model="gpt-4o"
    ) as gen:
        # Processing...
        gen.update_trace(tags=["summarization"])  # Adds to existing tags

When to use: For more control over span creation and nesting. Useful for complex workflows with multiple steps.

Update tags from anywhere within your traced code:

from abvdev import ABV

abv = ABV(
    api_key="sk-abv-...",
    host="https://app.abv.dev"
)

with abv.start_as_current_span(name="api-endpoint"):
    request_data = get_request()

    # Dynamically add tags based on request
    tags = ["api"]
    if request_data.get("beta_features"):
        tags.append("beta")
    if request_data.get("enterprise"):
        tags.append("enterprise")

    abv.update_current_trace(tags=tags)

    # ... processing ...

When to use: When tags depend on runtime conditions or need to be added from helper functions that don’t have direct span references.

Installation:

pip install abvdev

See Python SDK docs for complete reference.

JavaScript/TypeScript SDK

Setup
Context Manager
observe() Wrapper
Manual Spans

Install packages:

npm install @abvdev/tracing @abvdev/otel @opentelemetry/sdk-node dotenv

Add credentials to .env:

.env

ABV_API_KEY="sk-abv-..."
ABV_BASE_URL="https://app.abv.dev"  # or "https://eu.app.abv.dev" for EU

Create instrumentation.ts:

instrumentation.ts

import dotenv from "dotenv";
dotenv.config();

import { NodeSDK } from "@opentelemetry/sdk-node";
import { ABVSpanProcessor } from "@abvdev/otel";

const sdk = new NodeSDK({
  spanProcessors: [
    new ABVSpanProcessor({
      apiKey: process.env.ABV_API_KEY,
      baseUrl: process.env.ABV_BASE_URL,
      exportMode: "immediate",
      flushAt: 1,
      flushInterval: 1,
    })
  ],
});

sdk.start();

Using the context manager pattern:

import "./instrumentation"; // Must be first import

import {
  startActiveObservation,
  updateActiveTrace,
} from "@abvdev/tracing";

async function processQuery(query: string, useCache: boolean) {
  await startActiveObservation("query-processing", async (span) => {
    span.update({
      input: { query, useCache },
    });

    // Add tags based on configuration
    const tags = ["query-processing"];
    if (useCache) {
      tags.push("cached");
    }

    updateActiveTrace({ tags });

    // ... your processing logic ...
  });
}

processQuery("What is RAG?", true);

Wrapping existing functions:

import "./instrumentation"; // Must be first import
import { observe, updateActiveTrace } from "@abvdev/tracing";

async function fetchUserData(userId: string) {
  // Determine user tier
  const user = await getUser(userId);

  updateActiveTrace({
    tags: ["user-data", `tier:${user.tier}`, "api"],
  });

  // ... fetch data logic ...
  return userData;
}

// Wrap the function to trace it
const tracedFetchUserData = observe(fetchUserData, {
  name: "fetch-user-data",
});

// Use the wrapped version
const data = await tracedFetchUserData("user-123");

Creating spans manually:

import "./instrumentation"; // Must be first import
import { startObservation } from "@abvdev/tracing";

const span = startObservation("document-analysis", {
  input: { documentId: "doc-123" },
});

// Add tags to the trace
span.updateTrace({
  tags: ["document", "analysis", "production"],
});

// ... processing ...

span.update({ output: analysisResult }).end();

See JS/TS SDK docs for complete reference.

Best Practices

Keep Tags Simple and Consistent

Use lowercase, hyphenated strings for reliability and readability:Good examples:

rag, production, few-shot
error, beta, v2.1.0
document-processing, cache-hit

Avoid:

Spaces: "rate limit error" → use rate-limit-error
Mixed case: ProductionEnv → use production
Special characters: user@premium → use user-premium

Why it matters: Consistent formatting makes tags easier to filter, prevents duplicate categories, and ensures reliable UI behavior.

Establish Naming Conventions Early

Define and document tag patterns for your team before scaling:Option 1: Namespaced tags

tags = [
    "env:production",
    "version:2.1.0",
    "model:gpt-4o",
    "tier:premium"
]

Option 2: Simple tags

tags = [
    "production",
    "v2.1.0",
    "gpt4o",
    "premium"
]

Create a tag dictionary: Document your conventions in your team wiki or codebase:

# tags.py - Team tag conventions
ENVIRONMENTS = ["dev", "staging", "production"]
TECHNIQUES = ["rag", "few-shot", "chain-of-thought"]
ERROR_TYPES = ["rate-limit", "validation", "timeout"]

Why it matters: Prevents tag proliferation (prod vs production vs prd), ensures team-wide consistency, and makes onboarding easier.

Never Put PII in Tags

Tags are designed for categorization, not sensitive data. They appear in UI filters, exports, and analytics dashboards.❌ Don’t use:

# Bad - contains PII
tags = [
    "user:john@example.com",
    "customer:acme-corp",
    "account:123-45-6789"
]

✅ Instead use:

# Good - anonymous identifiers
tags = [
    "tier:premium",
    "region:us-east",
    "cohort:enterprise-2024"
]

# Use metadata for identifiers with proper access controls
metadata = {
    "user_id": "usr_abc123",  # Internal ID, not email
    "tenant_id": "tenant_xyz",
    "account_hash": hash(account_number)
}

Why it matters: Tags are visible to all project members and appear in exported data. PII in tags creates compliance risks and potential data leaks. Use metadata with appropriate access controls for sensitive identifiers.

Combine Tags with Metadata Strategically

Use both features together for maximum flexibility:Tags: Simple categories for filtering

tags = ["production", "rag", "error"]

Use when you need to:

Filter traces quickly in the UI
Create alerts on specific categories
Group metrics by common dimensions

Metadata: Detailed attributes for analysis

metadata = {
    "model_version": "gpt-4o-2024-08-06",
    "prompt_tokens": 1523,
    "chunks_retrieved": 5,
    "cache_hit": True,
    "retrieval_latency_ms": 245
}

Use when you need to:

Store structured data for custom queries
Track numeric metrics (tokens, latency, cost)
Include detailed context for debugging

Example combining both:

@observe()
def process_rag_query(query: str, user_tier: str):
    # Tags for filtering and categorization
    abv.update_current_trace(
        tags=["rag", "production", f"tier:{user_tier}"],
        metadata={
            "query_length": len(query),
            "retrieval_method": "vector-search",
            "embedding_model": "text-embedding-3-large",
            "chunks_count": 5
        }
    )

Why it matters: Tags give you fast filtering. Metadata gives you deep analysis. Together they provide both speed and depth.

Tag Dynamically Based on Runtime Conditions

Add tags based on execution paths, not just static configuration:

@observe()
def process_request(user_input: str):
    tags = ["api-request"]

    # Add tags based on validation
    if not is_valid_input(user_input):
        tags.append("validation-error")

    # Add tags based on execution path
    if needs_rag(user_input):
        tags.append("rag")
        tags.append("retrieval-heavy")
    else:
        tags.append("direct-completion")

    # Add tags based on performance
    start = time.time()
    result = execute(user_input)
    duration = time.time() - start

    if duration > 5.0:
        tags.append("slow-response")

    abv.update_current_trace(tags=tags)
    return result

Why it matters: Dynamic tagging captures what actually happened during execution, making it easier to identify patterns, debug issues, and optimize performance.

Limit the Number of Tags Per Trace

Keep tag lists focused and meaningful:Recommended: 3-7 tags per traceGood - focused and actionable:

tags = ["production", "rag", "gpt-4o", "premium-tier"]

Too many - loses signal in noise:

tags = [
    "production", "api", "v2", "rag", "gpt-4o",
    "cached", "premium", "us-east", "web",
    "document", "summarization", "high-priority"
]  # 12 tags - too many!

Strategy: Ask “Will I actually filter by this?” If not, put it in metadata instead.Why it matters: Too many tags make the UI cluttered and filtering less effective. Focus on tags you’ll actually use for filtering, grouping, or alerting.

Tags vs Metadata vs Environments

Choosing the right feature for your use case:

Feature	Best For	Example Use Cases
Tags	Simple categorization, filtering, experiments	`rag`, `production`, `beta`, `v2.1.0`, `few-shot`
Metadata	Structured data, detailed attributes, analytics	`{"tenant_id": "acme", "user_tier": "premium", "prompt_tokens": 1523}`
Environments	Formal separation with access controls	Development, Staging, Production projects

When to use multiple features together:

Tag with environment (production) AND use dedicated ABV Environments for formal separation
Tag with experiment variant (prompt:v2) AND include detailed metadata ({"variant_id": "abc123", "assignment_ts": "2024-01-15T10:30:00Z"})
Tag with feature (rag) AND include metadata about the RAG implementation ({"chunks": 5, "embedding_model": "text-embedding-ada-002"})

Metadata

Add structured key-value data to traces for detailed filtering and analytics

Environments

Separate development, staging, and production with dedicated projects and access controls

Sessions

Group related traces by user journey or workflow for end-to-end visibility

Trace IDs

Track requests across distributed services with custom trace IDs

Getting Started

Basic Features

LLM Gateway

Guardrails

Evaluations

Prompt Management

Cookbook

SDKs

Platform

Support

Tags

How Tags Work

Why Use Tags?

Implementation Guide

Best Practices

Tags vs Metadata vs Environments

Metadata

Environments

Sessions

Trace IDs

​How Tags Work

​Why Use Tags?

​Implementation Guide

​Best Practices

​Tags vs Metadata vs Environments

​Related Features

Metadata

Environments

Sessions

Trace IDs

How Tags Work

Why Use Tags?

Implementation Guide

Best Practices

Tags vs Metadata vs Environments

Related Features