langchain-highsnr 0.2.0

LangChain integration for HighSNR — compress documents to a token budget, keeping the highest-signal content.

langchain-highsnr

LangChain integration for HighSNR — compress documents to a token budget, keeping the highest-signal content.

pip install langchain-highsnr

Get an API key at console.high-snr.com. Requires Python 3.9+ and langchain-core>=0.3.0.

---

What it does

HighSNR selects the most informative chunks from a document and discards the rest, staying within a token budget. Compression is deterministic, privacy-first, and sub-second for most documents.

Class	Position in pipeline	Use case
HighSNRDocumentTransformer	Before embedding	Compress raw docs before indexing
HighSNRDocumentCompressor	After retrieval	Compress retrieved chunks before LLM

---

Usage

HighSNRDocumentTransformer — compress before indexing

from langchain_highsnr import HighSNRDocumentTransformer
from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings

transformer = HighSNRDocumentTransformer(
    api_key="snr-...",       # or set HIGHSNR_API_KEY env var
    max_output_tokens=800,   # token budget per document
    context_hint="clinical trial methodology",  # optional topic hint
)

compressed = transformer.transform_documents(raw_docs)
vectorstore = FAISS.from_documents(compressed, OpenAIEmbeddings())

Parameter	Default	Description
api_key	None	Falls back to HIGHSNR_API_KEY env var
max_output_tokens	1000	Token budget per document
include_boundaries	True	Keep first and last chunk
context_hint	None	Topic/query to bias chunk selection

---

HighSNRDocumentCompressor — compress after retrieval

The user's query is automatically used as the selection hint.

from langchain_highsnr import HighSNRDocumentCompressor
from langchain.retrievers import ContextualCompressionRetriever

compressor = HighSNRDocumentCompressor(
    api_key="snr-...",
    max_output_tokens=2000,
)

retriever = ContextualCompressionRetriever(
    base_compressor=compressor,
    base_retriever=vectorstore.as_retriever(search_kwargs={"k": 20}),
)

docs = retriever.invoke("what is the main finding?")

Parameter	Default	Description
api_key	None	Falls back to HIGHSNR_API_KEY env var
max_output_tokens	2000	Token budget across all chunks
include_boundaries	False	Keep first/last chunk
group_by_source	True	Group chunks by metadata["source"] — one API call per source document (benchmark-validated)

---

Benchmarks

Evaluated on LongBench v1 with GPT-4o (n=200 per dataset). At 80% budget with hint:

• HotpotQA: F1 70.96 — exceeds full-context GPT-4o (69.71)
• Qasper: F1 45.21 — 96% of full-context GPT-4o (47.22)

Full results, scripts, and reproduction instructions: github.com/HighSNRInc/highsnr-benchmarks

---

Environment variables

Variable	Description
HIGHSNR_API_KEY	API key — alternative to passing api_key in the constructor
HIGHSNR_API_URL	Override the API base URL (default: https://api.high-snr.com)

---

Links

• API console & free tier: console.high-snr.com
• Homepage: high-snr.com
• Benchmarks: github.com/HighSNRInc/highsnr-benchmarks
• Support: [e-mail]