What does provenance-aware priority handling do?

It prevents untrusted RAG, tool, and assistant content from gaining protected priority solely through imperative or safety-critical wording.

Does LLMSlim prevent prompt injection?

No. It mitigates compression-induced instruction elevation and applications still need defense in depth.

Does LLMSlim require remote calls?

The default extractive strategy runs locally. Rewrite and hybrid strategies use a caller-supplied provider.

Studio

Inspect the shape of context.

Run live extractive compression with the same LLMSlim Python package shipped from this repository. Output and token metrics come from the returned CompressionResult.

Configure

Choose context, then make the target yours.

Live LLMSlim engine

Sample library

RAG documentRetrieved policy with a malicious-looking irrelevant instruction.

ContextRoleStrategy

Target retention

Any whole number from 10 to 90.

Advanced

Maximum chunk sizeOptional LLMSlim max_chunk_tokens: 32–4000.

Provider paths stay protected. Rewrite and hybrid require a caller-supplied provider and are not exposed by the public Studio.

Input

Context editor

Run for real token count

Role: ragCtrl / Cmd + Enter to run

Run live extractive compression to view the exact output and token metrics returned by LLMSlim.

Live LLMSlim Python engine. Output and token metrics are returned by the same repository package used by the public API. Diff is derived from the visible input and output.

Live engine boundary

Python execution, with the right limits.

Set the correct context role when you know the provenance. RAG, tool, and assistant text are treated differently from system and developer text in the v0.4.0 safety boundary.

Run extractive live

The public Studio runs the offline extractive path through its same-origin Python function.

Bring a provider for rewrite

Rewrite and hybrid require a caller-supplied provider and are not exposed by the public endpoint.