Youssef

clef-compactor

Query-aware RAG context compaction with Cloudflare Clef: keep the evidence, cut noisy retrieval context.

Code on GitHub PyPI

Language
Python
Licence
Apache-2.0
Last push
16 Nov 2026
Stars
2

What it is

Query-aware RAG context compaction with Cloudflare Clef. Score an entire retrieval batch against the query, retain the chunks that earn their place, and send a smaller, auditable context to your LLM.

Quick start · Watch the demo · How it works · Benchmarks · Kaggle notebook

How it works

Retrieval can be noisy. If a retriever returns 12 chunks, your LLM normally reads, and you pay for, all 12, even when several are off-topic. clef-compactor asks Clef whether each chunk is needed for the user’s query, ranks the answers, then fills your token budget with the strongest evidence.

Results

What it does Why it matters
Keeps text verbatim Citations and provenance remain intact, nothing is summarized or rewritten.
Explains every cut Dropped chunks are marked irrelevant or budget_exhausted.
Batches requests One Cloudflare call scores up to 64 chunks.
Fits your stack Sync and async clients, CLI, OpenAI-compatible handler, LangChain, and LlamaIndex adapters.

Try it

pip install clef-compactor

# PowerShell
$env:CLEF_ACCOUNT_ID = "your_account_id"
$env:CLEF_API_TOKEN = "your_api_token"

# bash / zsh
export CLEF_ACCOUNT_ID=your_account_id
export CLEF_API_TOKEN=your_api_token