relay-llm 0.3.0

Multi-provider LLM batch prediction library

relay

A unified, provider-agnostic Python library for submitting, managing, and downloading results from large-scale LLM batch prediction jobs across Anthropic, OpenAI, Google, and XAI.

Installation

# Minimal
pip install relay

# All providers + full feature set
pip install relay[all]

# Individual providers
pip install relay[anthropic]
pip install relay[openai]
pip install relay[google]
pip install relay[xai]

Quickstart

import asyncio
from relay import BatchClient, BatchRequest, BatchConfig, Message

requests = [
    BatchRequest(id=f"req-{i}", messages=[Message(role="user", content=f"Summarize: {text}")])
    for i, text in enumerate(my_texts)
]

config = BatchConfig(provider="anthropic", model="claude-opus-4-5")

async def main():
    async with BatchClient() as client:
        job = await client.submit(requests, config)
        results = await client.wait_and_download(job.id)
    return results

results = asyncio.run(main())

Features

• Provider-agnostic API — switch providers with a single config change
• Resilient job state — jobs survive process crashes; state is persisted to SQLite
• Smart caching — identical (prompt, model, params) tuples are never sent twice
• Cost estimation — token count and cost estimate before every submission
• Budget controls — configurable warn, confirm, and hard-limit thresholds
• Fan-out — send the same request to multiple providers simultaneously
• Multiple export formats — JSONL, CSV, Parquet, HuggingFace Dataset
• Terminal dashboard — live TUI for monitoring jobs, costs, and cache
• Web dashboard — lightweight FastAPI UI for remote monitoring
• Observability — Prometheus, OpenTelemetry, and JSONL metrics exporters

Provider Compatibility

Provider	Model family	Batch API	Streaming	Fan-out
Anthropic	Claude 3/4	Yes	No	Yes
OpenAI	GPT-4o, o-series	Yes	No	Yes
Google	Gemini 1.5/2.0	Yes	No	Yes
XAI	Grok	Yes	No	Yes

CLI

# Submit a batch job
relay submit requests.jsonl --provider anthropic --model claude-opus-4-5

# Submit and wait for results
relay run requests.jsonl --provider openai --model gpt-4o --output results.jsonl

# Dry-run cost estimate
relay estimate requests.jsonl --provider anthropic --model claude-opus-4-5

# List and monitor jobs
relay jobs list
relay jobs status <job-id>

# Export results
relay export <job-id> --format parquet --output results/

# Cache management
relay cache stats
relay cache vacuum

# Spending report
relay costs today

Documentation

See the spec.md for the full technical specification and API reference.