llmtrim 0.1.11

Static, deterministic LLM prompt/payload compression that cuts input tokens 30-90% with zero extra model calls.

llmtrim-uniffi

UniFFI bindings over llmtrim-core: one Rust definition, idiomatic in-process bindings for Python, Ruby, Swift and Kotlin. The compression runs natively in the caller's process (no server, no async).

API

A deliberately flat surface over the engine:

fn compress(
    input: String,                 // a provider-shaped request body (JSON)
    provider: Option<Provider>,    // OpenAi | Anthropic | Google, or None to auto-detect
    preset: Option<String>,        // "aggressive" | "agent" | "code" | "rag" | "safe" | …
                                   // None = config from the environment / config file
) -> Result<CompressOutput, LlmtrimError>

CompressOutput carries the compressed request_json, the resolved provider/model, the tokenizer label/exactness, and the before/after/frozen input-token counts. Embedders that need the full rehydration plan or per-stage reports should depend on llmtrim-core directly in Rust.

Python

# Build a self-contained wheel (cdylib + generated glue):
crates/llmtrim-uniffi/scripts/build-wheel.sh --release
pip install target/wheels/llmtrim-*.whl

import llmtrim, json

req = json.dumps({"model": "gpt-4o",
                  "messages": [{"role": "user", "content": "…"}]})
out = llmtrim.compress(req, llmtrim.Provider.OPEN_AI, "aggressive")
print(out.input_tokens_before, "->", out.input_tokens_after)
# send out.request_json to the provider

Why build-wheel.sh and not plain maturin build: maturin's bindings = "uniffi" auto-packaging is sensitive to the maturin↔uniffi version pair. With maturin 1.14 + uniffi 0.31 it builds the native library into the wheel but omits the generated Python glue (empty package __init__.py). The script runs maturin, then injects the freshly generated bindings and repacks the wheel with valid RECORD hashes. Remove it once the auto path packages cleanly.

Ruby / Swift / Kotlin

All targets generate from the same built library, no extra Rust. The generated glue is a build artifact (its checksums are pinned to the library ABI), so it is regenerated per release rather than committed:

crates/llmtrim-uniffi/scripts/generate-bindings.sh out/   # python, ruby, swift, kotlin

Generation needs an unstripped library. Library-mode uniffi-bindgen reads metadata symbols from the cdylib, but the workspace release profile sets strip = true. The script therefore generates from the (unstripped) debug build; the native library you ship can be a stripped cargo build --release -p llmtrim-uniffi cdylib; the glue loads it by name.

Ruby (verified). This is the raw generated binding (module LlmtrimFfi), for a source build with libllmtrim_ffi.so on the load path. The published gem aliases it to Llmtrim (require "llmtrim" then Llmtrim.compress(...)); see packaging/ruby.

require_relative "llmtrim_ffi"
require "json"
out = LlmtrimFfi.compress(
  JSON.generate({model: "gpt-4o", messages: [{role: "user", content: "…"}]}),
  LlmtrimFfi::Provider::OPEN_AI, "aggressive")
puts "#{out.input_tokens_before} -> #{out.input_tokens_after}"

Swift emits llmtrim_ffi.swift + an FFI header and modulemap; Kotlin emits uniffi/.../llmtrim_ffi.kt (which loads the cdylib via JNA). CI compiles and runs a smoke for both: Swift on macOS (swiftc against the modulemap), Kotlin on a JVM (kotlinc + JNA), so a binding break is caught in all four languages (see tests/swift, tests/kotlin and the bindings* jobs in .github/workflows/ci.yml).

Publishable packages

Each ships the compiled engine bundled, so consumers need no Rust toolchain:

Target	Build	Package	Verified
Python (PyPI)	scripts/build-wheel.sh	wheel	locally
Ruby (gem)	scripts/build-gem.sh	packaging/ruby/	locally
Kotlin/JVM (Maven)	scripts/build-maven.sh	packaging/kotlin/	locally
Swift (SwiftPM)	scripts/build-xcframework.sh	packaging/swift/	macOS CI only

Each packaging/<lang>/README.md has the usage + publish details.