Adding Columns

generating the insights you need in tabular column types tabular supports two column types deterministic data fields are pulled directly from the data source with no language model processing these columns often load by default in new projects to characterize the results set, and can also be added via add column pane generative columns (denoted by the ai icon) use reliant’s ai model to extract specific information from source text, and are created via natural language prompts the user can enter any analytical prompt regarding their source document and receive precisely extracted data for each source in the table, forming a core part of tabular’s analysis capability let’s explore how to use both column types in your tabular projects adding a column after creating a project docid\ x6 8ly2ldfy0pt7nxsfco , tabular auto generates a results table with several columns depending on the project type and recipe used below is an output of an literature review (pubmed) search generated from the input clinical trial publications on alk positive cancers in the last 5 years by default, tabular will generate several default columns that give the user a sense of their result set for pubmed searches, default columns include the title, authors, journals, date, and publication types , with values pulled directly from each pubmed abstract to add new columns, in the upper right of the table this will open the column edit interface , which allows the user to review all active columns in the project, browse / add suggested columns , and add custom ai columns suggested columns are a curated list of columns likely to be relevant for the given project type (ex literature review vs clinical trial project) suggested columns are categorized by topic (ex overview, population, intervention, etc ), and are easily searchable at the top of the pane to add a suggested column, simply click on the column title in the add column pane this will open the prompt details window allowing the user to review the specific prompt used to generate the column users are highly encouraged to modify suggested column prompts to suit their specific extraction goals, and better understand prompting best practices here is an example of the population summary suggested column to create a custom column , simply click the button in the add column pane see below for prompting tips and examples once you are satisfied with the prompt for your research needs, click run extraction types of extraction output when creating a column, the user is able to select the output type desired tabular columns can return several types of output free response this output type will return concise, open ended text that best addresses the specified prompt this response type is highly flexible, and allows the user to instruct the model to respond in a specific way here are some output formatting examples specify categorical response list specify a list of potential categories from which you would like the model to select the most appropriate response which of the following conditions are studied in this paper b cell non hodgkin lymphoma (b nhl), follicular lymphoma (fl), or diffuse large b cell lymphoma (dlbcl), other (specify)? which of the following best describes the primary intervention discussed in the study dietary change or restriction, exercise adjustment, sleep habit adjustment, meditative or psychotherapy practice (ex mindfulness, talk therapy, etc )? specify data format return the primary efficacy endpoint duration in months in the following format x x months, endpoint name return the sex distribution of the enrolled population in this study, formatted as # male, # female if sex is not specified, return # sex not specified prompting for rationale does this paper provide evidence supporting a relationship between coffee and mental health? return yes or no, followed by a concise summary of the relevant supporting evidence yes / no this output type should be used only when asking yes or no questions the model will only return ‘yes’, ‘no’, or ‘unclear’ to prompting questions this output type is excellent for creating complex and precise filters to narrow results, critical to many literature review workflows yes / no column prompt examples is the primary intervention of this study a monoclonal antibody (mab)? is this paper a study of the pharmacokinetics (pk) of a molecule or drug? blank (fill in yourself) this output type allows users to create a named column of blank cells in their project for note taking and manual labeling purposes this output type has no associated prompt column generation tips more detail is often better! just like a human, the language model will search and synthesize information more precisely when it has more context on the request there is no restriction on generative prompt column length use plain language with specific context unlike when using traditional keyword search engines like google, it is beneficial to interact with the language model using prompts that mimic conversational language specificity is critical to extraction performance, and alignment with user goals give the model the same context you would give a human colleague here is strong and a weak example strong column example “is this product indicated for use in pediatric patients with renal cell carcinoma?” this example is strong because it specifies the context of the ask (indicated for use in a specific population), as well as the therapeutic area of interest weak column example “pediatric patient approval” this example is weak, as it does not specify that the user is asking about the currently indicated populations of the product as opposed to future populations under investigation this prompt also does not specify the indication of interest examples increase prompt efficacy just like a human coworker, reliant's model often performs better when given examples of the output you are looking for here are some sample prompts how many patients are included in this study? provide any descriptors of the patient population included (ex " 52 immunocompromised patients") (scientific literature project) what is the study type of this publication (ex observational study, meta analysis, modeling study, etc ) (scientific literature project) tabular can generate outputs in specific format if asked! specify example outputs, and tell the model you are looking for a specific format in the prompt here are some formatting prompts generate a reference for this article in the following format alinam, s , timms, j a , dillon, r , et al (publication year) (journal name) (scientific literature project) what is the post treatment or post operative qmax at each specified outcome time periods? use the following format 10 2 ml/s, 1 mo; 13 5 ml/s, 6mo; 24 5 ml/s, 12 mo (scientific literature project) iterate with different phrasing to continually imrove extraction we find that users regularly have more success crafting effective custom prompts after playing around with the tool don’t hesitate to try a few column options side by side to compare prompt efficacy! “n/a” result indicates lack of information the ai will return n/a if it cannot find information relevant to the question asked when asking a yes or no question, an n/a result is distinct from a no a no result indicates the ai has found information that is counter to the prompt, whereas n/a means no information was found to support nor deny the claim example generative ai columns describe the dosing for this product in detail, making sure to highlight if a site of care is mentioned (ex hospital, clinic, icu equipped facility) what clinical test measures are mentioned in the trial exclusion criteria? include specific numerical information (ex blood pressure > 140/90 mmhg) do not include age related exclusions list all drug names or molecules ending in "mab" that are listed as interventions in this trial does this abstract discuss wearable sensors or electronic data collection for hypertension patients? identify names of all clinical trials captured following this format trial name and nct code, comma separated (example "opal broaden, nct01877668") provide your answer as a simplified and concise list with no additional information