Trajbl
Benchmarks

Trajbl beat LLMLingua2 on quality-per-token.

The current public-facing evidence comes from controlled public benchmark suites such as LongBench-style multi-hop QA and WixQA-style support workloads. The fair interpretation: Trajbl is a post-retrieval compressor, not a full retriever.

Benchmark comparison showing Trajbl versus Microsoft LLMLingua2 with quality and context token metrics
Open the graphic for full-resolution detail.
+72%relative F1 lift vs LLMLingua2controlled long-context benchmark, similar token budget
~88%context token reductionmeasured against full-context RAG reference
68%RAG quality retainedwhile sending a much smaller context packet
0GPU/model dependencyTrajbl compression is CPU-first and model-free
Highest-quality branch

Full context is the quality ceiling. Trajbl keeps much of it with far less context.

In the long-context benchmark signal, Trajbl retained 68% of the full-context reference quality while cutting context tokens by about 88%. LLMLingua2 retained about 40% quality at a similar token budget.

Trajbl vs LLMLingua2Higher quality, lower tokens is better
Full RAG referenceQuality
0.587 F1Tokens
10,593
TrajblQuality
0.400 F1Tokens
1,313
LLMLingua2Quality
0.233 F1Tokens
1,405
Long-context comparison

Relative quality and token efficiency.

Higher quality is better. Lower context size is better. The headline is not raw retrieval accuracy; it is compression quality at a much smaller context budget.

MethodQualityQuality retainedMean context tokensToken reduction
Full RAG reference0.587 F1100%10,5930%
Trajbl0.400 F168%1,313~88%
LLMLingua20.233 F140%1,405~87%
Compression baselines

Trajbl also beat stronger LongLLMLingua-style baselines.

This is the clearest public-facing chart: Trajbl stays model-free and CPU-first while outperforming model-based compression baselines in the measured long-context comparison.

Trajbl vs LLMLingua2Higher quality, lower tokens is better
Full RAG referenceQuality
0.587 F1Tokens
10,593
TrajblQuality
0.400 F1Tokens
1,313
LLMLingua2Quality
0.233 F1Tokens
1,405
Against stronger compression baselinesObjective F1 comparison
Trajbl
0.398 F1
LLMLingua2 q-aware
0.238 F1
LongLLMLingua-7B
0.197 F1
Cost correction

~$0.26 actual LLM estimate

The old run summary showed $6.23 because the harness added a fixed $0.01 per judge row. The QA calls cost about $0.233 and the estimated judge cost is about $0.023. Future runs now log measured judge cost.

Next validation

Infra benchmark required

Before broad market claims, we will measure cold start, mean/p95 latency, RAM, CPU, throughput, and cost per 1,000 queries by stage: retrieval, BGE, Trajbl, LLMLingua2, LLM, and judge.

CurveMemory research appendix

Routing metrics remain promising, but they are no longer the main product headline.

CurveMemory is kept as a research track for geometric routing. The immediate product focus is Trajbl as a compressor for existing RAG and agent stacks.

CurveMemory routing latencyLogarithmic scale / lower is better
CurveMemory
0.11 ms
Semantic Router TF-IDF
1.27 ms
Semantic Router MiniLM
8.29 ms
FAISS Flat + MiniLM
6.14 ms
FAISS HNSW + MiniLM
5.89 ms
CurveMemory RAM deltaLogarithmic scale / lower is better
CurveMemory
0.02 MB
Semantic Router TF-IDF
6.16 MB
Semantic Router MiniLM
330.38 MB
FAISS Flat + MiniLM
195.7 MB
FAISS HNSW + MiniLM
197.95 MB
Accuracy vs total routing latencyUpper-left represents the strongest efficiency zone
60%70%80%90%CurveMemory: 81.8% at 0.109 msCurveMemorySemantic Router TF-IDF: 64.6% at 1.268 msSR TF-IDFSemantic Router MiniLM: 92.2% at 8.29 msSR MiniLMFAISS Flat + MiniLM: 93.1% at 6.14 msFAISS FlatFAISS HNSW + MiniLM: 93.1% at 5.89 msFAISS HNSWTotal routing latency (ms, log scale)
Routing index size across corpus scale10k to 250k documents / logarithmic vertical scale
10k50k100k250kCurveMemory: 0.36 MBCurveMemory: 1.08 MBCurveMemory: 1.81 MBCurveMemory: 2.91 MBFAISS HNSW: 16.02 MBFAISS HNSW: 80.12 MBFAISS HNSW: 160.25 MBFAISS HNSW: 400.62 MBFAISS Flat: 14.65 MBFAISS Flat: 73.24 MBFAISS Flat: 146.48 MBFAISS Flat: 366.21 MB
CurveMemoryFAISS HNSWFAISS Flat
CurveMemory measured data

Research routing methods at a glance.

MethodAccuracyTotal latencyRAM deltaIndex
CurveMemory81.8%0.109 ms0.02 MB55.5 KB
Semantic Router TF-IDF64.6%1.268 ms6.16 MBNot reported
Semantic Router MiniLM92.2%8.29 ms330.38 MB675 KB
FAISS Flat + MiniLM93.1%6.14 ms195.7 MB675 KB
FAISS HNSW + MiniLM93.1%5.89 ms197.95 MB738.3 KB
Scale data

CurveMemory research scale measurements.

DocumentsCurve latencyHNSW latencyCurve RAMHNSW RAMCurve index
10,0000.020 ms0.109 ms1.53 MB19.33 MB0.36 MB
50,0000.012 ms0.169 ms3.67 MB80.47 MB1.08 MB
100,0000.013 ms0.203 ms5.71 MB161.00 MB1.81 MB
250,0000.043 ms0.345 ms10.30 MB400.01 MB2.91 MB
Benchmark workstation

Intel Core i9-12900KF

16 cores / 24 logical processors / 32 GB RAM / Windows 11 Pro 64-bit

CurveMemory routing and CPU baseline measurements were executed on CPU. The installed RTX 3080 was not used for those routing measurements.

CPU12900KFLocal benchmark