Spaces:
Sleeping
Reproducibility Statement
This document describes the assumptions, constraints, and known sources of variation affecting the reproducibility of the DocVision system.
Hardware Assumptions
- CPU-only execution environment
- No GPU required
- Minimum 8 GB RAM recommended
- Stable internet connectivity required for external API access
- Tested on cloud-based Linux and local desktop environments
Runtime Estimates
Approximate end-to-end runtime per document (single page):
| Pipeline Stage | Avg Time (seconds) |
|---|---|
| OCR (LlamaParse) | 3.0 – 5.0 |
| Vision LLM Inference | 1.5 – 3.5 |
| Logo Detection (CPU) | < 10 |
| Total (Full Pipeline) | 14.5 – 18.0 |
Multi-page PDFs process only the first page by default, keeping runtime bounded.
Actual latency may vary depending on API load and network conditions.
Random Seed Handling
- Temperature for Vision LLM inference is fixed at
0.1 - Experiments declare a fixed seed value (
seed: 42) for documentation purposes - No stochastic sampling is intentionally introduced in the pipeline
- Deterministic preprocessing steps (PDF rendering, resizing, hashing)
Known Sources of Nondeterminism
Despite fixed configuration, some nondeterminism remains due to:
- Cloud-hosted Vision LLM inference
- OCR service variability across identical inputs
- Network latency fluctuations
- Concurrent execution order of asynchronous tasks
These factors may cause minor variations in extracted fields or reasoning text, but document classification remains stable in most cases.
Cost Considerations
- OCR and Vision inference rely on external APIs (LlamaParse, OpenRouter)
- Cost scales linearly with number of processed documents
- No GPU compute cost is incurred
- Logo detection and preprocessing run locally on CPU
- Caching mechanisms reduce redundant API calls for repeated inputs
Users should be aware of API usage limits and associated costs when processing large batches of documents.
Summary
DocVision prioritizes reproducibility through deterministic preprocessing, fixed inference parameters, and declarative experiment definitions. Remaining sources of nondeterminism stem primarily from external AI services rather than internal system design.