llmslim Files

Patch release for LLMSlim Core: framework adapters and better PyPI metadata. Compression behaviour is unchanged from 0.8.0.

:::bash
pip install --upgrade "llmslim[langchain]"    # or [llamaindex] (Python 3.10+) / [haystack]

Added

• LangChain: llmslim.integrations.langchain.LLMSlimCompressor, a BaseDocumentCompressor for ContextualCompressionRetriever (where you'd use LLMLinguaCompressor). Install llmslim[langchain].
• LlamaIndex: llmslim.integrations.llama_index.LLMSlimPostprocessor, a BaseNodePostprocessor (where you'd use LongLLMLinguaPostprocessor). Install llmslim[llamaindex]; requires Python 3.10+.
• Haystack: llmslim.integrations.haystack.LLMSlimDocumentCompressor, a component with run(documents, query). Install llmslim[haystack].
• What all three keep: document order, ids, scores and metadata. Documents that compress to nothing are dropped, and original and kept token counts are recorded in metadata.
• Options: target_ratio (default 0.25), relevance ("lexical" / "semantic" / a callable), selection, context_role, additional_compress_kwargs.
• Tested with:
• langchain-core 0.3.0 and 1.6.9
• llama-index-core 0.12.0 and 0.14.25
• haystack-ai 2.0.0, 2.21.0 and 3.3.0

Guides with runnable examples are in docs/integrations/.

Changed

• PyPI metadata:
• The summary now says what LLMSlim does.
• New keywords for prompt and context compression and the supported frameworks.
• Python 3.13, 3 :: Only and Typing :: Typed classifiers.
• llmslim[all] includes the three framework extras.

Benchmarks (unchanged, measured on 0.8.0): https://www.llmslim.app/benchmarks · methodology discussion: https://github.com/Thanatos9404/llmslim/discussions/36

Full changelog: CHANGELOG.md (github.com)