Context Optimization API

Context compression for LLM pipelines

HighSNR compresses long documents to fit your token budget before they reach your LLM — cutting costs without sacrificing answer quality.

✕ No AI involved 🔒 0 data retention = Same input → same output

Where it fits

LLM calls

Compress before you send

Pass a long document and a token budget. Get back only the chunks that matter. Fewer tokens sent — lower cost, faster responses, less hallucination from noise.

RAG Memory & Embeddings

Less vectors, less noise

Before embedding a large corpus, compress documents first. Fewer, higher-quality chunks mean less storage, faster retrieval, and less noise in your vector store.

How it works

✕

No AI inside

No model, no black box, no randomness. You know exactly what it does.

=

Deterministic

Same document, same budget, same output. Every time. No model drift, no surprises in production.

🔒

Zero retention

Your documents are never stored, logged, or used for training. Only counters and metadata are kept.

⚡

Fast

Sub-second for most documents. No local model download, no GPU required.

Benchmark

LongBench v1 · GPT-4o · n=200 per dataset · QA F1 score

Evaluated across two multi-hop and single-hop QA datasets. Higher is better.

HotpotQA

Config	50%	60%	70%	80%	Full doc
No hint	65.29	66.34	68.08	70.70	—
With query hint	67.28	68.02	69.95	70.96	—
Full context (baseline)	—	—	—	—	69.71

At 80% budget, HighSNR beats full-context F1.

Actual token ratio (output / input) — HotpotQA

Target	Mean	Median	Min	Max
50%	55.9%	55.4%	41.6%	71.7%
60%	67.9%	67.3%	55.0%	83.9%
70%	79.8%	79.1%	69.5%	99.9%
80%	91.4%	90.8%	81.1%	100.0%

Qasper

Config	50%	60%	70%	80%	Full doc
No hint	35.51	38.16	41.36	45.37	—
With query hint	39.87	40.76	42.97	45.21	—
Full context (baseline)	—	—	—	—	47.22

At 80% budget, HighSNR retains 96% of full-context F1 on scientific QA.

Actual token ratio (output / input) — Qasper

Target	Mean	Median	Min	Max
50%	54.7%	54.4%	37.5%	69.5%
60%	66.4%	66.2%	47.3%	79.7%
70%	78.0%	77.6%	69.2%	92.0%
80%	89.9%	89.5%	79.4%	100.0%

Actual ratios exceed the target because HighSNR never cuts a chunk mid-sentence. Chunks are selected whole — if the next chunk would exceed the budget it is skipped, so the output lands just below (not at) the target. Short documents where a single chunk spans the full budget pull the mean slightly above the target percentage.

Latency

Live API calls · 0.5 vCPU / 1 GB · n=3,200

< 5k tokens

770 ms median

mean 777 ms

5k – 10k tokens

1,102 ms median

mean 1,142 ms

10k – 20k tokens

1,792 ms median

mean 1,833 ms

API

One endpoint. Pass your document and a token budget. Get back the most relevant chunks.

POST /v1/optimize

curl https://api.high-snr.com/v1/optimize \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "document": "your long document text...",
    "budget": { "value": 2000 },
    "context_hint": "what is the main finding?"
  }'

Response

{
  "selected_chunks": [
    "Most relevant passage from your document...",
    "Second most relevant passage..."
  ],
  "selected_chunk_indices": [2, 5]
}

document

The full text to compress. Plain string.

budget.value

Max tokens in the output. Integer token budget for the selected chunks.

context_hint

Optional query string. Biases selection toward relevant chunks.