omneval-sdk 0.0.64

Python SDK for Omneval LLM tracing

Omneval Python SDK

Python SDK for Omneval LLM tracing and evaluation.

Installation

pip install omneval-sdk

Quick Start

import omneval_sdk

# Configure the SDK at application startup
omneval_sdk.configure(
    endpoint="http://localhost:4318",  # Omneval ingest endpoint
    api_key="oev_proj_your_api_key",
)

# Decorate functions to trace them
@omneval_sdk.trace
def call_llm(prompt):
    span = omneval_sdk.get_active_span()
    omneval_sdk.set_model(span, "gpt-4o")
    omneval_sdk.set_input(span, prompt)
    omneval_sdk.set_output(span, "Hello!")
    return "done"

result = call_llm("What is 2+2?")

API Reference

omneval_sdk.configure(endpoint, api_key)

Configure the OTLP HTTP exporter pointing at the Omneval ingest API. Must be called once at startup.

@omneval_sdk.trace

Decorator that wraps a function in an OTel span sent to Omneval. Nested decorated functions produce correctly linked parent-child span trees.

omneval_sdk.set_input(span, input_value)

Attach input data to a span as the omneval.input attribute.

omneval_sdk.set_output(span, output_value)

Attach output data to a span as the omneval.output attribute.

omneval_sdk.set_model(span, model)

Attach the model name to a span as the gen_ai.request.model attribute.

omneval_sdk.set_tokens(span, input_tokens, output_tokens)

Attach token counts to a span as gen_ai.usage.input_tokens and gen_ai.usage.output_tokens.

omneval_sdk.get_active_span()

Return the currently active span from the context variable. Returns None outside a decorated function.

omneval_sdk.OmnevalClient(base_url, api_key, project_id=None)

HTTP client for prompt fetching and manual score writes.

The project_id is automatically extracted from the API key suffix (e.g. oev_proj_my-project → my-project). Pass project_id explicitly to override this behavior.

client = omneval_sdk.OmnevalClient("http://localhost:8080", "oev_proj_my-project")
prompt = client.get_prompt("greeting", "production")
client.write_score("span-id-123", "helpfulness", 0.8, "Good reasoning")

client.write_score(span_id, eval_name, value, reasoning="")

Submit a manual score for a span. The project_id from the client configuration is automatically included in the request.

Args:

• span_id (str, required): The span ID to score.
• eval_name (str): Name of the evaluation (e.g. "helpfulness", "correctness").
• value (float): Numeric score value (typically 0.0–1.0).
• reasoning (str, optional): Human-readable reasoning for the score.

Returns: None on success (201 Created).

Raises:

• ValueError: If span_id is empty.
• requests.HTTPError: If the server returns an error (e.g. 400 Bad Request, 500 Internal Server Error).

client.write_score(
    span_id="span-abc123",
    eval_name="correctness",
    value=0.9,
    reasoning="The answer correctly identified the entity",
)

Exports

All key helper functions are exported from the top-level omneval_sdk package:

from omneval_sdk import (
    trace,           # @trace decorator
    configure,       # configure endpoint
    OmnevalClient,   # HTTP client
    set_input,       # set input on span
    set_output,      # set output on span
    set_model,       # set model on span
    set_tokens,      # set token counts on span
    get_active_span, # get current span
)