HighSNR
Less tokens, less noise — compress docs and RAG context
Details
- Categories
- AI
- Use Cases
- AI AgentsWorkflow Automation
- Target Audience
- DevelopersData ScientistsBackend Developers
- Platforms
- API
Discovery signals
How AI and people discover HighSNR on PeerPush
About HighSNR
HighSNR is a context compression/denoising API that removes low-signal passages from long documents and RAG retrieval results before they reach your LLM. Send a document and a token budget, get back only the high-signal chunks. Fully deterministic, zero data retention, no AI model in the processing pipeline — runs fast on commodity CPUs. Benchmarked on LongBench v1 with GPT-4o (200 samples): at 90% budget, beats full-context F1 on HotpotQA (71.57 vs 69.71) and retains 97.9% of full-context quality on Qasper. Budget is accurate end-to-end. Free tier: 2M tokens, 14 days, no card required. LangChain integration available via pip. Self-hosted option in progress for teams that need data to never leave their network.
Product Updates (1)
HighSNR v2 is out!
v2 is out with better signal selection, chunks input, and stronger benchmarks. The v2 engine is significantly more accurate at identifying high-signal passages, especially when you pass a query hint. Updated benchmarks (Claude Sonnet 4.5, n=200): - HotpotQA: beats full-document quality at 40–60% budget - Qasper: 97% of full-document quality at 50% of the tokens - Beats random selection by 5–15 F1 points across all budgets Holds across both GPT-4o and Sonnet 4.5, i.e., it's not model-specific. Free tier unchanged: 2M tokens, 14 days, no card.
Comments (0)
No comments yet. Be the first to share your thoughts!
Reviews (0)
No reviews yet. Be the first to rate this product!
Comments (0)
No comments yet. Be the first to share your thoughts!