Skip to main content

Pipeline & Throughput Benchmarks

End-to-End Pipeline Latency

StageTargetActualStatusPer-Page
Upload<500ms0.20s meanPASS~40ms
Processing5-30s7.71s meanPASS~1.54s
Search hybrid (no rerank)<50ms19.6ms mean of 5PASSN/A
Search dense-only<50ms19.2ms mean of 5PASSN/A
Search sparse-only<50ms8.8ms mean of 5PASSN/A
Reranking (20 docs)<2s1,469.67ms mean of 3PASSN/A
Total E2E5-35s26.93s wall clockPASS0.90s
Why can processing still be slow?

Processing time is dominated by document parsing, embedding, Qdrant upsert work, and the post-embedding ModernBERT ONNX classification callback. Some documents report near-zero poll duration because they completed before the polling request observed them.

Classification Benchmarks (Post-Embedding)

Classification runs after all embedding batches complete. It reads chunk embeddings from Qdrant, selects representative chunks via centroid + MMR, and scores them with ModernBERT ONNX.

Banking documents (5 pages, 11-17 chunks)

DocumentChunksCategories from live document metadata
bank_of_america_2024.pdf12financial reporting, risk management, corporate governance, regulatory compliance, Ledger
citigroup_2024.pdf11financial reporting, risk management, corporate governance
goldman_sachs_2024.pdf11financial reporting, risk management, corporate governance, business operations, market analysis
jpmorgan_chase_2024.pdf12financial reporting, corporate governance, regulatory compliance, Ledger
morgan_stanley_2024.pdf13financial reporting, risk management, corporate governance, market analysis, sustainability
wells_fargo_2024.pdf17financial reporting, risk management, corporate governance, regulatory compliance

Gold classification benchmark

MetricValue
Inline path accuracy, prefilter off5/6 (83.3%)
Post-embedding path accuracy, prefilter off5/6 (83.3%)
Median latency, prefilter off3.257s
Median latency, prefilter on0.521s
<8s per-document claimConfirmed
Classification variance

The gold benchmark missed q1-earnings-trading-update, where expected categories were financial reporting and market analysis, but the classifier returned financial reporting and risk management.

Page Throughput Analysis

MetricValue
Total Pages30
Wall Clock Time26.93s
Pages per Minute66.85
Seconds per Page0.90s
Avg Doc Latency7.91s

Per-Document Latency

DocumentUploadProcessingTotalChunks
bank_of_america_2024.pdf0.51s18.11s18.62s12
citigroup_2024.pdf0.51s18.11s18.62s11
goldman_sachs_2024.pdf0.05s0.00s0.05s11
jpmorgan_chase_2024.pdf0.05s2.01s2.05s12
morgan_stanley_2024.pdf0.04s0.00s0.05s13
wells_fargo_2024.pdf0.03s8.04s8.07s17

Embedding Stage Breakdown

StageMeanP50P95Count
NER0.5063s0.4685s0.7137s6
Dense embedding15.9442s14.8100s21.7145s6
Sparse embedding0.0055s0.0040s0.0087s6
Qdrant upsert0.0192s0.0105s0.0523s6
Total task16.8855s15.6760s22.8473s6

Resource Utilization

ResourceUsage
Total Documents7
Total Pages30 banking pages + 1 OCR image
Total Chunks78
Memory Usage Before Run81%
CPU Usage Before Run20%
Queue Status Before RunIDLE

Throughput Summary

MetricValue
Documents/minute13.37
Pages/minute66.85
Chunks/minute169.36
Search queries/second, hybrid~51.0
Search queries/second, dense-only~52.1
Search queries/second, sparse-only~113.6

Full Latency Reference

OperationMeanMedianRangeNotes
Upload0.20s0.05s0.03-0.51sBanking corpus upload
Processing (5 pages)7.71s5.02s0.00-18.11sPoll-observed duration
Search (hybrid, no rerank)19.6ms16ms16-27msHistorical 5-run mean, alpha=0.7
Search (dense-only)19.2ms17ms16-26ms5-run mean, alpha=1.0
Search (sparse-only)8.8ms9ms7-11ms5-run mean, alpha=0.0
Reranking (20 candidates)1,469.67ms1,468ms1,435-1,506ms3-run mean
OCR image processing8.08s8.08s8.08sSingle image rerun
Health check1,016.25ms1,016.25ms1,016.25msSingle check