async-batch-llm 0.20.0

Provider-neutral asyncio toolkit for concurrent LLM calls with retries, coordinated rate limits, bounded streaming, resumable checkpoints, and deadlines.

async-batch-llm

Run independent LLM calls concurrently with production-grade retries, coordinated rate-limit cooldowns, bounded input buffering, resumable checkpoints, deadlines, and complete token accounting.

The execution pipeline is provider-neutral: wrap your existing async client or use the built-in OpenAI-compatible, Gemini, or PydanticAI conveniences. Use it when you need results during the current workflow; latency-tolerant jobs may be better suited to a provider's native batch API.

Documentation · Getting started · Examples · Changelog

Quick start

Install the OpenAI and terminal-progress extras, then set OPENAI_API_KEY:

pip install 'async-batch-llm[openai,progress]'
export OPENAI_API_KEY='...'

import asyncio
from async_batch_llm import llm, process_prompts

async def main():
    batch = await process_prompts(
        llm("openai:gpt-4o-mini"), ["Summarize A", "Summarize B"],
        concurrency=10, progress=True,
    )
    print(batch.summary())

asyncio.run(main())

Run the credential-free embedded-application demo, or open the no-key notebook in Colab.

Other provider extras are gemini, openrouter, deepseek, and pydantic-ai; progress installs tqdm. The core package has no provider SDK dependency.

Use your existing async client

from async_batch_llm import (
    ArtifactIdentity,
    CallOutcome,
    CallableStrategy,
    ProcessorConfig,
    process_stream,
)


async def invoke(prompt, *, attempt, timeout, state):
    response = await existing_client.generate(prompt, timeout=timeout)
    return CallOutcome(
        response.text,
        token_usage=response.usage,
        metadata={"route": response.route},
    )


strategy = CallableStrategy(
    invoke,
    identity=ArtifactIdentity(provider="my-gateway", model="summary-route"),
)
config = ProcessorConfig(
    concurrency=32,
    max_queue_size=128,
    max_result_queue_size=64,
)

async for result in process_stream(strategy, database_prompt_source(), config=config):
    await save_result(result)

CallableStrategy is an adapter to the same execution path used by built-in strategies—not a second runtime. It adds bounded input/output handoff, concurrency admission, coordinated cooldowns, LLM-aware retries, per-item retry state, deadlines, checkpoint/replay, accounting, and observers around one existing async operation. See Use Your Existing Async Client.

Built-in providers and result handling

llm("provider:model") covers openai:, gemini:, openrouter:, and deepseek:; keyword arguments forward to the model constructor (e.g. llm("deepseek:deepseek-v4-flash", thinking=False, max_connections=150)). For custom clients, cached models, or custom strategies, use the explicit two-object form — OpenAIStrategy(OpenAIModel.from_api_key("gpt-4o-mini")) — described in the provider guides.

Pass (item_id, prompt) pairs to control IDs, or (item_id, prompt, context) triples to carry application data into each result. Collected results remain in completion order by default. Pass preserve_order=True, or call batch.in_input_order(), when stable submission order is required.

For incremental handling, stream results while a bounded work queue applies backpressure to the producer:

from async_batch_llm import ProcessorConfig, process_stream

config = ProcessorConfig(
    max_workers=50,
    max_queue_size=200,
    max_result_queue_size=100,
)

async for item in process_stream(strategy, huge_prompt_source, config=config):
    await save(item)  # completion order

max_queue_size bounds accepted input waiting for workers; max_result_queue_size bounds completed results waiting for the consumer. Both default to unbounded. process_prompts() retains every result by design.

Choose an execution surface

Need	API
Collect a finite run	process_prompts()
Handle results incrementally	process_stream()
Execute one resilient request	call() / call_result()
Share limits across service requests	LLMCallPool (LLMGateway compatibility alias)
Customize queueing and lifecycle	ParallelBatchProcessor

Batch, streaming, single-call, and shared-call execution share the same retry, timing, provider-admission, and token-accounting pipeline. See the single-call and shared-call guide and core API for the lower-level surfaces.

What the operational layer adds

Capability	Behavior
Error-aware retries	Separate budgets for content/transport failures and rate limits
Coordinated cooldowns	One worker's rate limit pauses the shared execution scope
Bounded streaming	Lazy sources and slow result consumers apply backpressure independently
Durable resume	Versioned JSONL checkpoints replay only compatible prior results
Guardrails	End-to-end item deadlines, batch deadlines, and category-based fail-fast
Accounting	Attempt timing and tokens include retries and failed provider calls
Observability	Typed lifecycle events, metrics, middleware, and progress callbacks

Why not just use gather?

A semaphore plus asyncio.gather() is enough when all you need is a concurrency cap. It does not provide coordinated 429 cooldowns, validation-aware retries, lazy producer backpressure, checkpoint-before-publication durability, or token accounting for failed attempts. return_exceptions=True also leaves application code to interpret exception objects mixed into the result list.

Use gather() for a small script when those operational guarantees do not matter. Use a provider's native batch API when delayed results are acceptable and its current pricing or throughput is a better fit. See the scenario-based comparison for Bespoke Curator, gateways, native batch APIs, and workflow engines.

A dated benchmark snapshot

In a June 10, 2026 GSM8K benchmark using a pre-release v0.12-era build:

• Thirty serial calls took 39–65 seconds; bounded worker pools completed them in 2.1–4.2 seconds on the uncapped providers.
• At equal concurrency over 1,000 prompts, the framework processed 72 items/s on DeepSeek and 108 items/s on Gemini 3.1, compared with 58 and 55 items/s for the benchmark's semaphore pool.
• The full 1,319-item run exposed retries, model escalations, permanent errors, token use, and cost/latency/accuracy tradeoffs in one result model.

These figures are historical evidence, not current provider guarantees. See the methodology, model IDs, pricing snapshot, and complete tables.

Production checkpoints and guardrails

The complete production resume example is runnable. The core configuration looks like this once your strategy and prompts are defined:

from pathlib import Path

from async_batch_llm import (
    AbortMode,
    ArtifactIdentity,
    GuardrailConfig,
    JsonlArtifactStore,
    ProcessorConfig,
    ResumePolicy,
    process_prompts,
)

store = JsonlArtifactStore(
    "runs/invoice-extraction.jsonl",
    identity=ArtifactIdentity(
        provider="openai",
        model="gpt-4o-mini",
        prompt_version="invoice-v4",
        parser_version="invoice-schema-v2",
        application_version="billing-pipeline-v7",
    ),
    fsync=True,
)
config = ProcessorConfig(
    max_workers=20,
    attempt_timeout=30,  # one provider attempt
    guardrails=GuardrailConfig(
        total_timeout_per_item=180,  # admission, waits, calls, and retries
        batch_timeout=3600,
        abort_on_error_categories=frozenset(
            {"authentication", "insufficient_balance"}
        ),
        abort_mode=AbortMode.DRAIN_ACTIVE,
    ),
)

batch = await process_prompts(
    strategy,
    prompts,
    config=config,
    artifact_store=store,
    resume=ResumePolicy.REUSE_SUCCESSES,
    preserve_order=True,
)

if batch.termination.kind != "completed":
    print("controlled stop:", batch.termination)
Path("summary.json").write_text(batch.to_json(), encoding="utf-8")

Important operational details:

• Check batch.termination. Batch deadlines and configured fail-fast stops return completed and collateral terminal results rather than disguising the controlled stop as an unexpected exception.
• Each newly executed terminal result is appended and flushed before it is returned or streamed. fsync=True requests stronger durability; the default is flush-only.
• JsonlArtifactStore serializes concurrent writes within one process. It does not claim cross-process append safety.
• Replay compatibility includes item ID, prompt, participating context, and the complete artifact identity—not merely item_id.
• REUSE_SUCCESSES reruns prior failures. REUSE_ALL also replays compatible terminal failures.
• Raw prompts and contexts are excluded by default. Outputs and metadata are included by default and may contain sensitive application data.
• Historical replay tokens remain on each result for audit, while live processor statistics exclude them from current-run consumption.

Read Results, Artifacts, and Resume and Deadlines and Fail-Fast Guardrails for schema compatibility, privacy controls, abort modes, and deadline details.

Provider-neutral execution

Built-in strategies cover:

• OpenAIStrategy, OpenRouterStrategy, and DeepSeekStrategy through the shared OpenAI-compatible model layer.
• GeminiStrategy, including structured response parsing and shared context caching.
• PydanticAIStrategy for PydanticAI agents and typed output.

Anthropic can be used through PydanticAI or CallableStrategy. Other OpenAI-compatible services can reuse the common model layer or be wrapped as an existing async client. Subclassing LLMCallStrategy remains available for more specialized integrations; built-in provider models are not required.

Model identifiers and service limits change independently of this package. Confirm current provider documentation when choosing a model, connection pool, or concurrency limit. See the custom strategy guide and OpenAI-compatible high-throughput guide.

Timing, retry, and ordering semantics

The Choosing Your Limits guide walks every limit below in decision order — from concurrency= through connection pools, admission, timeouts, deadlines, ramp, and cooldown — with a worked 10k-item sizing example.

• attempt_timeout limits one provider execution attempt (renamed from timeout_per_item in v0.20; the old name is a deprecated alias).
• GuardrailConfig.total_timeout_per_item limits the complete logical item, including coordinated cooldown, startup ramp, proactive rate limiting, provider-capacity admission, calls, retry cooldowns, and backoff.
• GuardrailConfig.batch_timeout starts when the processor run starts.
• A fail-fast category triggers only after an item reaches terminal failure; retryable intermediate attempts do not abort the batch.
• AbortMode.DRAIN_ACTIVE lets an in-progress provider call finish. AbortMode.CANCEL_ACTIVE cancels unfinished accepted work while preserving external caller-cancellation semantics.
• Collected and streamed results remain completion ordered by default. process_prompts(..., preserve_order=True) orders a collected batch by its stable submission index. Streaming intentionally remains completion ordered to avoid blocking behind a slow early item and buffering later results.

See the production checklist and bounded-work guide for queue, connection-pool, and lifecycle guidance.

Results, serialization, and accounting

BatchResult aggregates input, cached, output, and total tokens across retries, including usage recovered from failed attempts. Cost remains caller-supplied; the package does not bundle a provider price table:

cost = batch.estimated_cost(
    input_per_mtok=current_input_rate,
    output_per_mtok=current_output_rate,
    cached_token_rate=current_cache_rate,
)

WorkItemResult and BatchResult support strict, versioned JSON and JSONL serialization. Unsupported values raise instead of silently falling back to repr(). Dataclasses, Pydantic models, enums, dates, UUIDs, paths, tuples, and sets serialize to JSON-safe values; use an encoder/decoder pair when typed reconstruction is required. Exception descriptors never restore arbitrary classes or tracebacks.

See the artifact and serialization API and core API.

Testing without provider calls

Use the included fake strategies and MockAgent to exercise latency, rate limits, retryable failures, and terminal failures without spending API quota. The project test suite makes no live provider calls. See the testing guide.

Examples

Start with these runnable examples:

• Existing async application client, bounded streaming, and replay
• Production checkpoints and guardrails
• OpenAI batch processing
• Single calls and a shared call pool
• Validation-aware model escalation
• Custom embedding strategies

Browse the complete examples directory for Gemini, DeepSeek, OpenRouter, Anthropic, LangChain, caching, grounding, and benchmark walkthroughs.

Documentation

• Getting Started
• Compare Alternatives
• Production Checklist
• Troubleshooting and FAQ
• Results, Artifacts, and Resume
• Deadlines and Fail-Fast Guardrails
• Bounded Work and Backpressure
• API Reference

Contributing

Clone the repository and use its pinned development environment:

git clone https://github.com/geoff-davis/async-batch-llm.git
cd async-batch-llm
uv sync --all-extras
make ci

See the contributing guide or open an issue. For operational help, start with the troubleshooting guide.

License

MIT License. See LICENSE.