Methodology
AI LedParser uses the model layer not to hallucinate engineering parameters, but to read complex datasheets: finding relevant graphs, recognizing their structure, and preparing data for verifiable extraction. We do not disclose internal heuristics, verification thresholds, or post-processing details, but openly describe the overall system workflow, application boundaries, and output requirements.
Final numbers are not "guessed" by an AI model. The model helps parse the document and graph, while final LED characteristics are calculated by a separate deterministic computational layer.
At the core of the service is a hybrid AI pipeline: the model layer helps interpret PDF and visual graph structure, while the calculation layer handles reproducible numerical operations. First, the PDF is processed as a document: pages and regions containing relevant graphs are identified. Then scales, units of measurement, and working ranges are reconstructed for the selected graph. After that, the visual curve is converted into a numerical series. Based on this series, calculations suitable for engineering work are performed.
For practical calculation of luminous flux, efficacy, and lumen cost, use the luminous flux calculator.
Graphs in an LED datasheet define key LED performance characteristics under operating conditions: forward current and voltage (I-V curve), power, luminous flux, efficacy (lm/W), and thermal derating. The tool identifies these graphs in the PDF and converts them into numerical series, enabling parameter calculations at the required operating current rather than relying on average table values.
This allows the system to treat LED technical specifications as verifiable data: voltage–current (I–V) dependencies, flux and power changes at a given operating current, rather than fragmented numbers. Individual parameters such as color rendering index (CRI), CCT, and I-V curve shape are discussed in dedicated materials.
The numerical result is not produced in a single model pass, but through a sequence of separate, verifiable steps.
The service locates relevant pages and crops the graph area within the PDF.
Axes, scales, labels, and physical units are detected.
The visual curve line is converted into a set of points in real coordinate space.
Derived parameters are computed by a separate calculation layer, not generated by the model.
Accuracy is achieved not by promising a "smart model", but by strictly separating roles: AI helps read and structure the datasheet, but cannot fabricate curves, scales, or final parameters.
We deliberately did not entrust LLMs with "predicting" curve shapes or axis ranges. That shortcut looks simpler, but fails to provide the required 2% accuracy. The model layer does not replace measurement: numerical output only appears after controlled graph cropping, explicit axis calibration, and point extraction.
There is no single universal model in the pipeline entrusted with the entire result. Specialized LLM models handle narrow, segmented tasks and respond strictly in structured, deterministic formats suitable for subsequent verification and processing.
Final validation remains with the human engineer: they must confirm data extraction accuracy. If discrepancy exceeds 2%, the system disallows graph verification and suggests reporting an error for model fine-tuning.
The system performs well with datasheets having clear graphs, visible axis titles, and legible tick marks. The extraction method does not yet support multiple curves on a single graph, nor graphs with non-uniform grid steps or heavily cluttered, noisy backgrounds.
