Performance Benchmarks
No buzzwords, only measured numbers · all data comes from project test reports, verification welcome
Four sets of core measured data
From P2/P3 end-to-end tests, DeepDoc pipeline measurements and the pipeline test report
End-to-end full-flow processing
(full marks on the structure layer)
Dual-model collaborative layout analysis
(GPT-4o basis)
Parsing quality: how to read the three-layer scores
Every processed document receives a health-check sheet
443-page OFD document measured details
Structure layer (whether layout blocks, heading hierarchy and reading order are correct) —full marks; after combining the relation layer (table row-column relationships, context continuity) and the content layer (text, OCR accuracy), the overall score is 0.924. Table structure mis-detection was converged through dual-model collaboration 64%。
Three-Layer Evaluation System
Going live is only the beginning — continuous health checks keep it reliable
Corpus level: is the raw material good
Measures the quality of preprocessing itself, scored automatically each time a batch of documents is ingested.
- Parsing completeness
- Table structure preservation rate
- OCR accuracy
- Redaction coverage
- Permission hit rate
Retrieval level: is retrieval accurate
Quantifies retrieval effectiveness with standard question sets to locate recall problems.
- HitRate@K / Recall@K
- MRR (mean reciprocal rank)
- NDCG@K (ranking quality)
Generation level: are answers correct
Measures the faithfulness and relevance of AI answers, directly tied to business experience.
- Faithfulness (fidelity to the source)
- AnswerRelevancy (off-target answer rate)
- ContextPrecision (context precision)
How evaluation closes the loop: one thumbs-down, and the system finds the cause itself
After a user gives a thumbs-down to an answer, the system automatically traces back the knowledge points the answer cited and attributes the issue to a specific step — was semantics cut by chunking, content mistakenly deleted by cleaning, redaction overdone, or metadata missing? The located result flows back into the evaluation set automatically and is regression-verified automatically at the next strategy change; releases are blocked until metrics pass the regression gate.
Compression performance: numbers from our original craft
The compression engine is the platform's foundation and a direct cost item
Typical compression results
| Metric | Measured value | What it means for customers |
|---|---|---|
| Average compression ratio | 60% (up to 80%) | Storage costs cut by more than half |
| Processing speed | In-memory architecture, 3-5x faster | In-memory processing across the whole flow, no temporary-directory disk IO |
| Image quality gating | SSIM ≥ 0.92 / PSNR ≥ 30dB | Smaller does not mean blurrier — optimal parameters found automatically |
| Font deduplication | Saves 30-50% of font space | Automatic merging of redundant fonts in Office documents |
| Multimodal token savings | 85.9% (GPT-4o basis) | Feed compressed images to large models and inference bills drop substantially |
What numbers will your documents produce?
Book a test run with your real documents and we deliver the report