trulens-notebook-execution

TruLens Notebook Execution

Execute Jupyter notebooks, display progress to the user, and handle API key requirements.

When to Use This Skill

Use this skill when:

• Running TruLens quickstart or example notebooks
• Demonstrating TruLens functionality via notebooks
• Testing notebook examples end-to-end
• User asks to "run the notebook" or "execute the notebook"

Execution Method

Always use jupyter nbconvert --execute to run notebooks. This:

• Maintains state across cells (variables persist)
• Captures all output properly
• Handles async operations correctly
• Works with OTEL tracing

DO NOT try to run notebooks by:

• Extracting cells and running them individually in bash
• Using python -c with heredocs
• Running as a standalone Python script (loses notebook context)

Basic Execution Command

jupyter nbconvert --to notebook --execute --inplace <notebook_path>

Execution with Timeout (for long-running notebooks)

jupyter nbconvert --to notebook --execute --inplace \
  --ExecutePreprocessor.timeout=600 \
  <notebook_path>

Execution with Output to stdout

jupyter nbconvert --to notebook --execute --stdout <notebook_path>

Displaying Progress to User

When running a notebook, display section headers as each cell executes - NOT generic "BASH_OUTPUT" messages.

Step 1: Parse the Notebook Structure First

Before executing, read the notebook JSON to build a map of:

• Markdown headers (## Section Name)
• Which code cells belong to which section

import json

with open('notebook.ipynb') as f:
    nb = json.load(f)

sections = []
current_section = "Setup"

for i, cell in enumerate(nb['cells']):
    if cell['cell_type'] == 'markdown':
        source = ''.join(cell['source'])
        # Extract header
        for line in source.split('\n'):
            if line.startswith('## '):
                current_section = line.replace('## ', '').strip()
    sections.append((i, current_section))

Step 2: Display Section Headers During Execution

When checking output or between cell groups, display the section name:

=== Step 1: Create the Search Tool ===
[cell output here]

=== Step 2: Create the Deep Agent ===
[cell output here]

=== Step 3: Set Up TruLens Session ===
[cell output here]

Key Rule: Never Show "BASH_OUTPUT" to User

When polling for bash output during notebook execution:

• DO: Print the current section header before showing output
• DON'T: Just say "BASH_OUTPUT" or "checking output..."

Example Display Implementation

# When starting a section:
print(f"\n=== {section_name} ===")

# When showing cell output:
print(output)

# When section completes:
print("✓ Complete")

Progress Display Pattern

Running notebook: deep_agents_quickstart.ipynb

=== Step 1: Create the Search Tool ===
✓ Complete

=== Step 2: Create the Deep Agent ===
✓ Complete

=== Step 3: Set Up TruLens Session ===
Starting dashboard...
Dashboard started at http://localhost:8501
✓ Complete

=== Step 4: Define Agent GPA Feedback Functions ===
✓ Complete

=== Step 5: Instrument the Agent with TruGraph ===
✓ Complete

=== Step 6: Run and Evaluate ===
Running agent with question: "What is the weather in San Francisco?"
Agent response: "The weather in San Francisco is..."
Waiting for evaluation results...
✓ Evaluations complete

=== Results ===
Answer Relevance: 1.0
Tool Selection: 1.0
...

Handling API Keys

Critical: Check environment first, then prompt for keys ONE AT A TIME

Step 1: Check Environment

env | grep -E "OPENAI|TAVILY|ANTHROPIC" || echo "No API keys found"

Step 2: If Keys Not Found, Prompt Individually

When prompting for keys:

• Ask for ONE key at a time
• Use the key prefix as a hint in the option label (e.g., "sk-proj-..." for OpenAI)
• Let users paste directly - don't rely on complex "Other" field workflows

Example prompt pattern:

Question: "Paste your OPENAI_API_KEY:"
Header: "OpenAI"
Options: [{"label": "sk-proj-...", "description": "Paste your sk-... key"}]

The user will paste their actual key by selecting "Other" or the option itself will be replaced with their input.

Step 3: Set Keys When Running

OPENAI_API_KEY="sk-..." TAVILY_API_KEY="tvly-..." \
  jupyter nbconvert --execute ...

Common API Keys for TruLens Notebooks

Keeping the Dashboard Alive

Critical: The notebook execution process ends, killing any dashboard started within it.

After notebook execution completes, launch the dashboard separately using TruLens's run_dashboard() function.

Important: Database Location

The notebook writes its database to ./default.sqlite relative to the notebook's directory. The run_dashboard() function reads from ./default.sqlite relative to the current working directory.

This means you MUST cd to the notebook's directory before launching the dashboard.

Correct Pattern for Dashboard Persistence

cd /path/to/notebook/directory && \
python3 << 'EOF'
from trulens.core import TruSession
from trulens.dashboard import run_dashboard

session = TruSession()
run_dashboard(session)
EOF

Use run_in_background=true with the bash tool so the dashboard stays alive.

Why NOT to Use Native Streamlit Commands

DO NOT try to launch the dashboard with native streamlit commands like:

# WRONG - will connect to wrong/empty database!
streamlit run /path/to/trulens/src/dashboard/trulens/dashboard/main.py

This fails because:

1. Streamlit runs from the current working directory (likely repo root)
2. It looks for ./default.sqlite relative to that directory
3. The actual database is in the notebook's directory
4. Result: "No apps found" in the dashboard

Full Example Workflow

# Step 1: Execute notebook
OPENAI_API_KEY="sk-..." jupyter nbconvert --execute --inplace \
  /path/to/examples/notebook.ipynb

# Step 2: Launch persistent dashboard FROM THE NOTEBOOK'S DIRECTORY
cd /path/to/examples && \
python3 << 'EOF'
from trulens.core import TruSession
from trulens.dashboard import run_dashboard

session = TruSession()
run_dashboard(session)
EOF
# Use run_in_background=true for this command

The dashboard will output its URL (e.g., http://localhost:55872) and remain running until explicitly stopped.

Post-Execution

After notebook execution:

1. Show the leaderboard/results if available
2. Provide the dashboard URL if one was launched
3. Summarize what was evaluated (metrics used, scores achieved)

Troubleshooting

Fixing Invalid Notebook JSON

If you see Additional properties are not allowed ('id' was unexpected):

import json

with open('notebook.ipynb', 'r') as f:
    nb = json.load(f)

# Remove 'id' fields from cells (not valid in nbformat 4)
for cell in nb['cells']:
    if 'id' in cell:
        del cell['id']

with open('notebook.ipynb', 'w') as f:
    json.dump(nb, f, indent=1)

Example Workflow

User: "Run the deep agents quickstart notebook"

1. Read notebook to identify:
   - Section headers (for progress display)
   - Required API keys (OPENAI_API_KEY, TAVILY_API_KEY)

2. Check environment for existing keys:
   env | grep -E "OPENAI|TAVILY"

3. Prompt for missing keys (ONE AT A TIME):
   "Please provide your OPENAI_API_KEY:"
   [User enters key]

   "Please provide your TAVILY_API_KEY:"
   [User enters key]

4. Execute notebook, displaying section headers:

   === Step 1: Create the Search Tool ===
   ✓ Complete

   === Step 2: Create the Deep Agent ===
   ✓ Complete

   === Step 3: Set Up TruLens Session ===
   ✓ Complete

   === Step 4: Define Agent GPA Feedback Functions ===
   ✓ Complete

   === Step 5: Instrument the Agent ===
   ✓ Complete

   === Step 6: Run and Evaluate ===
   Running agent...
   Waiting for evaluation results...
   ✓ Complete

5. Launch dashboard in background FROM THE NOTEBOOK'S DIRECTORY:
   cd /path/to/notebook/directory && python3 -c "
   from trulens.core import TruSession
   from trulens.dashboard import run_dashboard
   session = TruSession()
   run_dashboard(session)
   "
   [run_in_background=true]

6. Display results summary:
   "✓ Notebook execution complete!

    Evaluation Results:
    - Answer Relevance: 1.0
    - Tool Selection: 1.0
    - Tool Calling: 1.0
    - Execution Efficiency: 0.33
    - Plan Quality: 1.0
    - Plan Adherence: 1.0

    Dashboard running at: http://localhost:8501
    (Dashboard will stay alive until you stop it)"

Integration with Other Skills

This skill works alongside:

• instrumentation/ - for understanding what's being traced
• evaluation-setup/ - for understanding feedback functions
• running-evaluations/ - for interpreting results