Python Client Reference
Phoenix Client Documentation
Python Example
View on GitHub
TypeScript Example
View on GitHub
Overview
Available in arize-phoenix-client 1.28.0+ (Python) and @arizeai/phoenix-client 2.0.0+ (TypeScript) Creating datasets from traces has been available in the Phoenix UI. Now, the Phoenix client libraries enable programmatic dataset creation from production traces with bidirectional links between dataset examples and their source spans. Query spans using filters, transform the data, and create datasets that maintain connections to the original traces.How It Works
The workflow involves three steps: query spans from traces using filters, transform the span data into dataset format, and create a dataset with span associations using thespan_id_key parameter (Python) or spanId field (TypeScript).
Querying Spans
Query spans by attributes such as model name or other span attributes.Transforming Data
Parse span attributes and extract the fields needed for your dataset examples. Span IDs must be preserved to maintain the link to source traces.Linking to Source Spans
Python: Usespan_id_key to specify which column contains span IDs.
TypeScript: Use spanId field in each example object.
Phoenix validates that span IDs exist and creates bidirectional links visible in the UI.
Complete Examples
See the full examples in the phoenix-client repositories:- Python:
create_dataset_from_spans_example.py - TypeScript:
create_dataset_from_spans.ts
Requirements
- Phoenix server running (local or hosted)
- Existing traces with LLM spans
- Python:
arize-phoenix-client1.28.0 or later - TypeScript:
@arizeai/phoenix-client2.0.0 or later