Prela SDK

Fire-and-forget trace ingestion SDK for TraceShield.

Installation

pip install prela

Quick Start

from prela import Prela

prela = Prela(
    api_key="ts_ingest_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
    endpoint="https://api.traceshield.example.com",
)

# Manual trace
prela.trace(
    input="What is my account balance?",
    output="I don't have access to account information.",
)

# OpenAI integration
from prela.integrations.openai import PrelaOpenAI
from openai import OpenAI

client = PrelaOpenAI(OpenAI(), prela)
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
)

# Before process exit
prela.shutdown()

Design

• Fire-and-forget: Never blocks the agent's response path
• Zero dependencies: Only requires httpx
• Batching: Buffers traces and sends every 500ms or 50 traces
• Resilient: Drops traces on failure, logs warnings, never raises

Telemetry

Send runtime events from your AI agent:

import prela

# Configure once
prela.configure_telemetry(
    api_key="ts-key-...",
    org_id="org-...",
    agent_id="agent-...",
)

# Track events throughout your code
prela.track("llm_call", {"model": "gpt-4", "tokens": 150, "latency_ms": 1200})
prela.track("policy_violation", {"rule": "R-001", "input": "..."}, severity="warning")
prela.track("error", {"message": "Timeout connecting to API"}, severity="error")

# Flush before exit
prela.flush()

Events are batched (default: every 5s or 10 events) and sent to TraceShield's telemetry API.

License

Private. All rights reserved.