justllm 0.3.2

Production LLM calls. Just the three lines. Reliability, native caching, and reversible context compression on by default.

justllm

Production LLM calls. Just the three lines.

from justllm import LLM

llm = LLM("anthropic/claude-opus-4-8")
reply = llm("Summarize this contract.")

Cross-provider fallback, native prompt caching, and reversible context compression are on by default. No config. The surface stays tiny on purpose — the moment you need a dozen knobs, that is what LiteLLM is for.

Why

The ecosystem split in two: feature-complete but heavy (LiteLLM, LangChain), or simple but feature-thin (aisuite, any-llm). Nobody ships the production layer behind a three-line front door. justllm is that middle.

The one number that makes it worth a switch: compressing the dynamic junk that bloats agent calls — tool outputs, logs, RAG dumps — cuts the input-token bill without touching your code. Measured here (gpt-4o token basis): 53% saved on a JSON API tool result, 97% on repetitive logs, with a safe no-op when compression wouldn't help. The engine is Headroom (PyPI: headroom-ai, content-aware and reversible); justllm applies it only to tool/retrieved content, never to your prompts. See benchmarks/.

Install

pip install 'justllm[all]'     # transport + structured output + compression

Or take only what you need: justllm[litellm] (real calls), justllm[structured] (extract()), justllm[compression] (Headroom). The bare pip install justllm gives you the API and the reliability layer; calls raise a clear error until a transport is installed.

Usage

# fallback chain + explicit knobs
llm = LLM(
    chain=["anthropic/claude-opus-4-8", "openai/gpt-5", "groq/llama-3.1-70b"],
    compress=True,     # reversible, dynamic-context only
    cache="prompt",    # "cache" never silently means semantic
)

# structured output — a validated Pydantic instance
from pydantic import BaseModel

class Invoice(BaseModel):
    vendor: str
    total: float

inv = llm.extract(Invoice, "Parse: Acme Corp billed $4,200")

# a minimal tool-calling agent (tool outputs are auto-compressed)
agent = llm.agent(system="You are a travel assistant.", max_steps=8)

@agent.tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return weather_api(city)

agent.run("What should I pack for Boston this weekend?")

# streaming
for chunk in llm.stream("Tell me a short story."):
    print(chunk, end="")

# async (acall / aextract)
reply = await llm.acall("Summarize this.")

# opt-in routing: short prompts -> cheap model, long -> strong (no extra call)
from justllm import Router
routed = LLM(router=Router(small="groq/llama-3.1-8b-instant", large="openai/gpt-4o"))

# cheap-first cascade: escalate to the strong model only when needed
from justllm import Cascade
smart = LLM(router=Cascade(small="groq/llama-3.1-8b-instant", large="openai/gpt-4o"))

# load prompts from files (no registry); only your {vars} are substituted
from justllm import prompts
prompt = prompts.load("summary", document=text)   # reads prompts/summary.txt

# ...or back the same seam with a registry (Langfuse, etc.)
prompts.set_loader(prompts.langfuse_loader(label="production"))
prompt = prompts.load("summary", document=text)   # now fetched from Langfuse

Optional OpenTelemetry tracing (pip install 'justllm[otel]') emits gen_ai.* spans with a per-call gen_ai.usage.cost — the dollar figure the OTel spec leaves out. No-op until you configure a collector.

Status

Alpha (0.3.0). Wiring is unit-tested with mocked providers (no network in CI), and the call paths are validated live against Ollama and Groq:

• Calls — sync llm("..."), async llm.acall(...), and llm.stream(...), all through LiteLLM and wrapped in cross-provider fallback.
• Structured output — llm.extract(Model, ...) / await llm.aextract(...) return a validated Pydantic instance (via instructor).
• Reliability — with_fallback / awith_fallback + RetryPolicy: retry-with-jitter on retryable errors only, one retry layer.
• Caching — Headroom's per-provider cache optimizer (Anthropic breakpoints; OpenAI/Google handled) plus an opt-in exact-match cache.
• Compression — compress over Headroom (tunable via CompressConfig); agent tool outputs are compressed automatically.
• Routing — opt-in Router (length-based) and Cascade (cheap-first, escalate only when needed); deterministic, no extra judge call.
• Prompts — prompts.load(name, **vars) with a pluggable seam: the default file loader, a hot-reloading file_loader (mtime-cached), and a Langfuse registry adapter; no registry built in.
• Observability — optional OpenTelemetry GenAI spans with the per-call gen_ai.usage.cost the spec omits; no-op unless [otel] is installed.
• Agent — a minimal tool-calling loop with a hard step cap.

Live behavior is still alpha — exercise it with your own keys (or local Ollama) via benchmarks/bench_e2e.py.

Benchmarks

pip install -e '.[benchmarks]'
python -m benchmarks.run

Measures token/cost savings from compression, the overhead the layer adds, and that fallback actually recovers provider failures. The suite runs even without the optional deps (using fallbacks), so it is never a hard blocker.

Contributing

Contributions are very welcome — the goal is to stay SOTA and easy to use at the same time. Start with CONTRIBUTING.md (especially the design principles that keep the surface small), see where things are headed in ROADMAP.md, and look for good first issue labels.

License

MIT