batchfactory 0.1.1

Composable, cache-aware batch processing pipelines for LLMs, APIs, and dataset generation.

BatchFactory

Composable, cache‑aware batch processing pipelines for LLMs, APIs, and dataset generation.

Status: alpha – expect breaking changes. Names and APIs will shift rapidly.

---

Install

pip install batchfactory   # coming soon: pip install -U batchfactory

---

Quick start

import batchfactory as bf

project = bf.CacheFolder("example1", 1, 0, 0)
broker = bf.brokers.ConcurrentLLMCallBroker(project("cache/llm_broker.jsonl"))

# Build a small graph that rewrites passages into short English poems

g = (
    bf.ReadMarkdownLinesOp("./data/*.txt", "keyword", directory_str_field="directory")
    | bf.ShuffleOp(42)
    | bf.TakeFirstNOp(3)
    | bf.GenerateLLMRequestOp(
        'Rewrite the passage from "{directory}" titled "{keyword}" as a four-line English poem.',
        model="gpt-4o-mini@openai",
    )
    | bf.ConcurrentLLMCallOp(project("cache/llm_call1.jsonl"), broker)
    | bf.ExtractResponseTextOp()
    | bf.SaveJsonlOp(project("out/poems.jsonl"),output_fields=["keyword","text","directory"])
    | bf.PrintTextOp()
)

g = g.compile()
g.resume()
g.pump(dispatch_broker=True, reset_input=True)

---

Core ideas (WIP)

• Batch‑centric: every op consumes/produces lists of Entry objects.
• Atomic Ops: small, single‑purpose units you compose with | or wire().
• Graph‑first: compile once, then resume() and pump() as external brokers finish work.
• Cache‑mindful: transparent on‑disk ledgers allow you to pause and restart without recomputation.
• Broker pattern: offload long‑running or external work (LLMs, search APIs, human labeling) to pluggable brokers.
• Loop‑aware: native For, While, and If controls let graphs loop through subgraphs cleanly—ideal for multi-agent or staged workflows.

---

Roadmap

• Stabilise op names
• Replace OpGraph executor with pluggable schedulers
• Public docs & more examples

---

© 2025 · MIT License