Sessions

Sessions provide a way to group related traces together, making it easier to understand multi-turn conversations, user interactions, or related operations.

What are Sessions?

A session represents a logical grouping of traces that belong together, such as:

Multi-turn conversations with a chatbot
User sessions across multiple requests
Related operations in a workflow
A/B test variants for comparison

Why Use Sessions?

Sessions help you:

Track conversations across multiple LLM interactions
Analyze user behavior over time
Debug issues by viewing all related traces together
Calculate metrics across multiple operations
Group experiments for comparison

Setting Session IDs

Using `using_session` Context Manager

The recommended way to set session IDs:

from openinference.instrumentation import using_session
from openai import OpenAI

client = OpenAI()

# All traces within this context belong to the same session
with using_session(session_id="user-123-conversation-1"):
    response1 = client.chat.completions.create(
        model="gpt-4",
        messages=[{"role": "user", "content": "Hello!"}]
    )
    
    response2 = client.chat.completions.create(
        model="gpt-4",
        messages=[
            {"role": "user", "content": "Hello!"},
            {"role": "assistant", "content": response1.choices[0].message.content},
            {"role": "user", "content": "What's the weather?"}
        ]
    )

All traces created within the using_session context will have the same session ID in their metadata.

Manual Session Metadata

You can also set session metadata manually on spans:

from opentelemetry import trace
from openinference.semconv.trace import SpanAttributes
import json

tracer = trace.get_tracer(__name__)

with tracer.start_as_current_span("chat-turn") as span:
    # Set session ID in metadata
    metadata = {
        "session_id": "user-123-session-456",
        "user_id": "user-123",
        "conversation_id": "conv-789"
    }
    
    span.set_attribute(
        SpanAttributes.METADATA,
        json.dumps(metadata)
    )

Session ID Strategies

User-Based
Conversation-Based
Time-Based
Workflow-Based

import uuid
from openinference.instrumentation import using_session

def handle_user_request(user_id: str, message: str):
    # Create session per user
    session_id = f"user-{user_id}"
    
    with using_session(session_id=session_id):
        return process_message(message)

import uuid
from openinference.instrumentation import using_session

class ChatSession:
    def __init__(self):
        self.session_id = str(uuid.uuid4())
        self.messages = []
    
    def send_message(self, message: str):
        with using_session(session_id=self.session_id):
            response = llm.generate(self.messages + [message])
            self.messages.append(message)
            self.messages.append(response)
            return response

# Each conversation gets unique session
chat = ChatSession()
chat.send_message("Hello!")
chat.send_message("How are you?") # Same session

from datetime import datetime
from openinference.instrumentation import using_session

def handle_request(user_id: str, message: str):
    # Session per hour per user
    hour = datetime.now().strftime("%Y-%m-%d-%H")
    session_id = f"{user_id}-{hour}"
    
    with using_session(session_id=session_id):
        return process_message(message)

import uuid
from openinference.instrumentation import using_session

def process_document(document_id: str):
    # Session for entire document processing workflow
    session_id = f"doc-{document_id}-{uuid.uuid4()}"
    
    with using_session(session_id=session_id):
        # All steps share same session
        chunks = chunk_document(document_id)
        embeddings = generate_embeddings(chunks)
        store_embeddings(embeddings)

Real-World Example: Chat Application

Here’s a complete example of using sessions in a chat application:

import uuid
import gradio as gr
from openinference.instrumentation import using_session
from opentelemetry import trace
from opentelemetry.trace import Status, StatusCode
from openinference.semconv.trace import SpanAttributes
from phoenix.otel import register
from openai import OpenAI

# Setup tracing
register(project_name="chatbot", auto_instrument=True)

client = OpenAI()
tracer = trace.get_tracer(__name__)

class ChatSession:
    def __init__(self, user_id: str):
        self.user_id = user_id
        self.session_id = str(uuid.uuid4())
        self.messages = []
    
    def chat(self, user_message: str) -> str:
        """Process a chat message within a session."""
        # Use session context
        with using_session(session_id=self.session_id):
            # Create a span for the entire chat turn
            with tracer.start_as_current_span(
                "chat-turn",
                attributes={
                    SpanAttributes.OPENINFERENCE_SPAN_KIND: "CHAIN",
                    SpanAttributes.INPUT_VALUE: user_message,
                    SpanAttributes.METADATA: json.dumps({
                        "user_id": self.user_id,
                        "turn_number": len(self.messages) // 2 + 1
                    })
                }
            ) as span:
                try:
                    # Add user message
                    self.messages.append({
                        "role": "user",
                        "content": user_message
                    })
                    
                    # Get LLM response (automatically traced)
                    response = client.chat.completions.create(
                        model="gpt-4",
                        messages=self.messages
                    )
                    
                    assistant_message = response.choices[0].message.content
                    self.messages.append({
                        "role": "assistant",
                        "content": assistant_message
                    })
                    
                    span.set_attribute(
                        SpanAttributes.OUTPUT_VALUE,
                        assistant_message
                    )
                    span.set_status(Status(StatusCode.OK))
                    
                    return assistant_message
                    
                except Exception as e:
                    span.record_exception(e)
                    span.set_status(Status(StatusCode.ERROR, str(e)))
                    raise

# Gradio interface
sessions = {}  # Store active sessions

def chat_interface(user_id: str, message: str, history):
    # Get or create session for user
    if user_id not in sessions:
        sessions[user_id] = ChatSession(user_id)
    
    response = sessions[user_id].chat(message)
    return response

if __name__ == "__main__":
    demo = gr.ChatInterface(
        fn=chat_interface,
        title="Phoenix Chat (with Sessions)"
    )
    demo.launch()

Session Metadata

Enrich sessions with additional metadata:

from openinference.instrumentation import using_session
from opentelemetry import trace
from openinference.semconv.trace import SpanAttributes
import json

tracer = trace.get_tracer(__name__)

with using_session(session_id="session-123"):
    with tracer.start_as_current_span("operation") as span:
        # Add rich session metadata
        metadata = {
            "session_id": "session-123",
            "user_id": "user-456",
            "user_email": "[email protected]",
            "session_start_time": "2024-01-15T10:00:00Z",
            "client_type": "web",
            "user_tier": "premium",
            "experiment_variant": "variant-a"
        }
        
        span.set_attribute(
            SpanAttributes.METADATA,
            json.dumps(metadata)
        )

Filtering by Session

In Phoenix UI

Open Phoenix at http://localhost:6006
Navigate to the Traces view
Use the search bar: metadata.session_id:"your-session-id"
View all traces for that session together

Using Phoenix Client

import phoenix as px

client = px.Client(endpoint="http://localhost:6006")

# Get all spans for a session
spans = client.get_spans(
    filter_condition='metadata.session_id == "session-123"',
    limit=1000
)

# Analyze session metrics
for span in spans:
    print(f"Span: {span.name}")
    print(f"  Duration: {span.latency_ms}ms")
    print(f"  Tokens: {span.attributes.get('llm.token_count.total', 0)}")

Session Analytics

Analyze sessions to understand user behavior:

import phoenix as px
import pandas as pd

client = px.Client()

# Get all traces
spans_df = client.get_spans(limit=10000)

# Extract session IDs from metadata
def extract_session_id(metadata):
    if pd.isna(metadata):
        return None
    try:
        import json
        meta = json.loads(metadata) if isinstance(metadata, str) else metadata
        return meta.get('session_id')
    except:
        return None

spans_df['session_id'] = spans_df['metadata'].apply(extract_session_id)

# Calculate session metrics
session_stats = spans_df.groupby('session_id').agg({
    'latency_ms': ['mean', 'sum', 'count'],
    'attributes.llm.token_count.total': 'sum'
}).round(2)

print("Session Statistics:")
print(session_stats)

# Find longest sessions
longest_sessions = session_stats.sort_values(
    ('latency_ms', 'count'),
    ascending=False
).head(10)

print("\nLongest Sessions (by turn count):")
print(longest_sessions)

Session Best Practices

Use Meaningful Session IDs

Include context in your session IDs:

# Good: Descriptive session IDs
session_id = f"user-{user_id}-conv-{conversation_id}"

# Avoid: Opaque session IDs without context
session_id = str(uuid.uuid4())

Add Rich Metadata

Include metadata that helps with filtering and analysis:

metadata = {
    "session_id": session_id,
    "user_id": user_id,
    "session_type": "support_chat",
    "started_at": datetime.now().isoformat()
}

Consistent Session Boundaries

Define clear start and end points for sessions:

# Start new session for each conversation
class Conversation:
    def __init__(self):
        self.session_id = str(uuid.uuid4())
        self.started_at = datetime.now()
    
    def end(self):
        # Log session completion
        logger.info(f"Session {self.session_id} ended")

Handle Session Expiration

Clean up old sessions:

from datetime import datetime, timedelta

class SessionManager:
    def __init__(self):
        self.sessions = {}
        self.session_timeout = timedelta(hours=1)
    
    def get_session(self, user_id: str):
        if user_id in self.sessions:
            session, last_active = self.sessions[user_id]
            if datetime.now() - last_active < self.session_timeout:
                self.sessions[user_id] = (session, datetime.now())
                return session
        
        # Create new session if expired or doesn't exist
        new_session = ChatSession(user_id)
        self.sessions[user_id] = (new_session, datetime.now())
        return new_session

Sessions vs Projects

Feature	Sessions	Projects
Scope	Groups related traces	Isolates entire applications
Duration	Short-lived (minutes to hours)	Long-lived (entire app lifetime)
Use Case	User conversations, workflows	Applications, environments
Filtering	Via metadata search	Top-level selector
Isolation	Logical grouping only	Complete data isolation

You can use both together:

from phoenix.otel import register
from openinference.instrumentation import using_session

# Set project for the application
register(project_name="production-chatbot", auto_instrument=True)

# Use sessions to group conversations
with using_session(session_id="user-123-conv-456"):
    # Traces go to "production-chatbot" project
    # And are grouped in session "user-123-conv-456"
    process_chat()

Get Started

Core Features

Tracing

Evaluation

Datasets & Experiments

Integrations

What are Sessions?

Why Use Sessions?

Setting Session IDs

Using `using_session` Context Manager

Manual Session Metadata

Session ID Strategies

Real-World Example: Chat Application

Session Metadata

Filtering by Session

In Phoenix UI

Using Phoenix Client

Session Analytics

Session Best Practices

Sessions vs Projects

Next Steps

Annotations

Cost Tracking

Build docs developers (and LLMs) love

Get Started

Core Features

Tracing

Evaluation

Datasets & Experiments

Integrations

​What are Sessions?

​Why Use Sessions?

​Setting Session IDs

​Using using_session Context Manager

​Manual Session Metadata

​Session ID Strategies

​Real-World Example: Chat Application

​Session Metadata

​Filtering by Session

​In Phoenix UI

​Using Phoenix Client

​Session Analytics

​Session Best Practices

​Sessions vs Projects

​Next Steps

Annotations

Cost Tracking

Build docs developers (and LLMs) love

What are Sessions?

Why Use Sessions?

Setting Session IDs

Using `using_session` Context Manager

Manual Session Metadata

Session ID Strategies

Real-World Example: Chat Application

Session Metadata

Filtering by Session

In Phoenix UI

Using Phoenix Client

Session Analytics

Session Best Practices

Sessions vs Projects

Next Steps