← /blog
· ACE Engineering#llm #backend #cost #prompt-compaction #tokens #gateway #eval #scorecard #research #managed-api-stack

Prompt Compaction v1: A Scorecard

An 827-example benchmark evaluating ACE's local ONNX LLMLingua-2 context compactor: achieving 43.2% input token reduction while maintaining 97.2% load-bearing information retention.

Prompt Compaction v1: A Scorecard

In context-heavy AI workloads—such as document Q&A, retrieval-augmented generation (RAG), and multi-turn agentic workflows—input tokens dominate API spend. In our introduction to Prompt Compaction, we outlined how intelligent, entropy-based token pruning strips low-information prose from incoming user messages before they leave the API gateway.

With Prompt Compaction v1, we conducted a comprehensive 827-prompt evaluation benchmark comparing our local, quantized ONNX transformer compressor (LLMLingua-2) against stopword heuristic baselines across 10 diverse task domains.

The results are definitive: at our default 0.50 keep-ratio, Prompt Compaction v1 delivers 43.2% net input token savings while retaining 97.2% of critical, load-bearing information. Compared to basic stopword filtering, contextual transformer scoring cuts key information loss by 2.44× ($p = 1 \times 10^{-6}$), unlocking dramatic cost reductions without compromising application reliability.


How Prompt Compaction v1 Works

Compaction operates transparently on the hot path of the API gateway. Before a request is dispatched to an upstream provider, the compressor evaluates prose informativeness while enforcing strict structural protections:

Incoming User Message
  │
  ├── 1. Structural Protection ──► Code Fences (```), JSON ({...}), & System Instructions
  │                                (Preserved byte-for-byte; never scored, never dropped)
  │
  ├── 2. Contextual Scoring  ──► Local ONNX Quantized Transformer (LLMLingua-2)
  │                                Scores context-aware informativeness per word
  │
  ├── 3. Top-K Token Retention ──► Retain top-scoring 50% of prose tokens
  │
  └── 4. Deterministic Rejoin  ──► Rejoin surviving prose with protected spans
                                  compacted message dispatched to provider

Zero-Risk Code & Schema Safety

Prompt Compaction v1 automatically partitions code blocks, JSON payloads, and system prompt instructions prior to scoring. Fenced code and structured schemas are reproduced byte-for-byte, guaranteeing that syntax structure is never corrupted during compression.


Empirical Benchmark Scorecard ($n=827$ Prompts)

We evaluated performance across three operational keep-ratios on an independent 827-prompt dataset ($n=777$ non-contaminated out-of-domain prompts).

  • Retention: Percentage of derived must-keep answer terms and critical numeric literals surviving rewrite.
  • Token Savings: Actual input token reduction measured via provider BPE tokenizers.
Keep Ratio Setting Compressor Engine Must-Keep Retention Real Token Savings Prompts Losing Key Information Statistical Significance
0.75 (Gentle) Stopword Heuristic 99.9% 20.3% 2 / 777 Baseline
0.75 (Gentle) Learned ONNX 99.8% 20.6% 4 / 777 Equivalent ($p = 0.69$)
0.50 (Default) Stopword Heuristic 92.1% 40.7% 95 / 777 Baseline
0.50 (Default) Learned ONNX 97.2% 43.2% 39 / 777 Decisive ($p = 1 \times 10^{-6}$)
0.33 (Aggressive) Stopword Heuristic 60.8% 58.5% 426 / 777 Baseline
0.33 (Aggressive) Learned ONNX 87.0% 60.5% 156 / 777 Decisive ($p < 1 \times 10^{-15}$)

Key Takeaways

  1. The 0.50 Ratio Sweet Spot: At ratio=0.50, the learned ONNX compactor achieves 97.2% retention while saving 43.2% of input tokens (a 1.76× compression factor).
  2. 2.44× Quality Advantage: On 90 distinct prompts where simple stopword filtering deleted critical answer terms, contextual transformer scoring correctly preserved them.
  3. Conservative Operating Boundaries: Compacting down to 0.33 causes substantial information degradation across all models and is not recommended for production.

Domain Performance & Adversarial Robustness

Performance Across Workload Domains (ratio=0.50)

Workload Domain Benchmark Source Heuristic Retention Learned ONNX Retention Quality Retention Delta
Dense Prose RAG rag-prose 95.3% 100.0% +4.7%
Factual QA & Retrieval qa-triviaqa 84.2% 98.3% +14.1%
Long-Context Search qa-longbench 87.0% 96.4% +9.4%
Multi-Hop Reasoning qa-hotpot 87.2% 96.1% +8.9%
Agent Tool Trajectories agent-trajectory 100.0% 100.0% 0.0%
Multi-Turn Chat History chat-history 100.0% 100.0% 0.0%

Context-aware token classification shines brightest on dense prose RAG and long-context retrieval (TriviaQA +14.1%, LongBench +9.4%, HotpotQA +8.9%). On agent tool logs and conversational histories, structural span protection ensures 100.0% retention of file paths, status codes, and UUIDs.


Latency & Architecture Efficiency

Running inference 100% locally on CPU using a quantized int8 110M parameter transformer model, Prompt Compaction v1 delivers microsecond-to-millisecond execution speeds:

Prompt Token Length Median Latency (p50) Latency Overhead vs Upstream LLM Call
$< 1,000$ tokens 45 ms ~0.03% overhead
$1,000 – 4,000$ tokens 457 ms ~0.30% overhead
$4,000 – 12,000$ tokens 2,744 ms ~1.80% overhead

Byte-Stable Prompt Cache Compatibility

Prompt Caching relies on exact byte prefix matching. ACE Prompt Compaction v1 applies compression per message, ensuring that recurring prompt prefixes compact to byte-identical strings every time. This guarantees complete compatibility with upstream provider prompt caches.


Business Impact & Enterprise Financials

By trimming 43.2% of unneeded context tokens from high-volume API streams, Prompt Compaction v1 converts unused prompt capacity into direct enterprise cost savings.

                  ┌────────────────────────────────────────────────────────┐
                  │          Input Token Cost Spectrum per 1B Tokens       │
                  └───────────────────────────┬────────────────────────────┘
                                              │
         ┌────────────────────────────────────┴───────────────────────────────────┐
         │                                                                        │
┌────────▼────────────────────────────────┐              ┌────────────────────────▼───────────────┐
│ Mid-Tier Pricing ($3.00 / MTok Input)   │              │ Flagship Pricing ($5.00 / MTok Input)  │
│ Uncompressed Cost: $3,000 / 1B Tokens   │              │ Uncompressed Cost: $5,000 / 1B Tokens  │
└────────────────┬────────────────────────┘              └────────────────┬───────────────────────┘
                 │                                                        │
                 │  Compacted Cost: $1,704 / 1B Tokens                     │  Compacted Cost: $2,840 / 1B Tokens
                 ▼                                                        ▼
   43.2% Direct Net Cost Reduction                         43.2% Direct Net Cost Reduction
   ($1,296 Savings per 1B Tokens)                          ($2,160 Savings per 1B Tokens)

Summary of Enterprise Benefits

  • Direct Financial Savings: Reduces input token billing by 43.2%, saving $1,296 to $2,160+ per 1 Billion input tokens on standard commercial model tiers.
  • Expanded Effective Context Window: Enables long-context workloads to fit comfortably inside standard context windows without triggering unexpected context truncation errors.
  • Enterprise-Grade Safety: Delivers 97.2% load-bearing information retention with 100% syntax protection for code and JSON payloads.

Looking Ahead

Prompt Compaction v1 proves that intelligent, context-aware token pruning can eliminate nearly half of incoming API prompt costs while preserving application quality. As context windows expand and multi-agent workflows proliferate, local context optimization will remain a vital layer of modern cloud API infrastructure.

To learn more about implementing Prompt Compaction or to explore our open evaluation harnesses, visit the ACE Fleet Documentation.


References

  1. Y. Pan, Z. Tan, et al. LLMLingua-2: Data Distillation for Efficient Prompt Compression. ACL 2024. arXiv:2403.12968
  2. H. Jiang, Q. Wu, et al. LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models. EMNLP 2023. arXiv:2310.05736
  3. K. Łajewska et al. (Amazon). Information Preservation in Context Compression for RAG. EMNLP 2025. arXiv:2503.19114