Source Viewer Tool & Reliant Score

tracking data lineage in tabular when using generative ai tools for sensitive life sciences applications, it is critical to understand where software responses come from, and if they are supported by source data reliant is built with data transparency, accuracy, and accountability at its core, and users always have access to source data via the source viewer tool source viewer tool when clicking any cell within a tabular result table, users can explore the sources used to generate cell output by clicking the panel icon at the top right of the table users can also right click a cell and select view sources this will open the source viewer pane tabular will identify where supporting evidence for each extraction was found in the source data users can review the highlighted text to rapidly validate tabular’s data extraction additionally, users always have the option to open the source in its original location by clicking the icon in the top right of the source viewer pane in this case, this button links to the abstract in pubmed the source viewer may contain sections that reflect the structure of the source type (ex abstracts may have methods, results, and discussion sections), but will be rendered similarly across the app pdf source viewer reliant’s pdf source viewer is the most advanced in the industry, allowing users to natively view each pdf file in app , and explore highlighted content where evidence supports extracted cell data when data is extracted from a chart or table, reliant’s pdf source viewer will highlight the appropriate figure when multiple pieces of information support the extracted data, the pdf source viewer allows users to quickly parse through each highlighted area of supporting evidence reliant score you’ll notice that at the top of the source viewer in the above examples, there is a green “high" reliant score rating for each result cell, tabular generates an extraction confidence score that measures how well the model believes it addressed the given field/prompt for default, raw text columns that extract data that is directly defined in the source, (ex fda label indications, clinical trial start date, abstract pmid, etc ) the confidence level will automatically be 100% this is because an ai language model is not used to generate the result strings for cells that are created via an ai summary or generative ai prompt (see adding columns docid\ dq0rcozhbvzhjmok8ch3k here for more information), the confidence of the response may vary based on the quality of relevant information in the source material, as well as the phrasing of the prompt itself when an extraction with lower confidence is detected, an orange orange icon will indicate that user review is suggested tabular’s confidence ratings are conservative by design the software will err on the side of caution to avoid overstating accuracy to users we recommend using the source viewer pane for lower confidence outputs to understand how tabular is extracting data for a given prompt occasionally with low confidence answers, tabular can fail to find relevant information to highlight this is a good indication that the cell extraction is not supported by robust data, and users should be cautious if using this data without manual verification