Skip to main content
Prompts are instructions that guide your AI Agent on how to process and analyze your data. This page will explain the basics of Prompts including defining output types, attaching files, and advanced usage like chaining Prompts together and performing complex math.

​
Output Types

Each Prompt has an output type. This type specifies the format of the response you expect. When creating a Prompt, Kolena will intelligently guess the output type based on your instructions, but you can also select the desired output type from the dropdown menu. Click to expand each output type for more details:
Text Prompts are the most flexible output type. They allow the Agent to generate free-form text responses based on the Prompt instructions.Text output can also be formatted using Markdown syntax. This is useful for creating structured documents with headings, lists, links, and other formatting elements.
Table Prompts enable structured data representation. It is useful for extracting and organizing information in a tabular format. When Table is selected, you must define the columns that the Agent should fill in. Accuracy will be highest when your Prompt includes clear instructions on what data should go into each column as well as instructions for when to create a row.You can use Table Prompts to directly extract tabular data from documents, such as a rent roll from a PDF, or to organize unstructured information into a table, such as gathering a list of all critical dates within a legal document. The number of rows returned by the Agent will depend on the content of the input document and your Prompt instructions.
Accuracy of Table Prompts will decrease with very large (20+) numbers of columns or when each row requires many paragraphs of text to be extracted. Recommendation: if the number of columns becomes too large, split your instruction into multiple Prompts.
Form Prompts return a set of predefined field/value pairs. Each field in the Form corresponds to a specific piece of information that the Agent is expected to extract or infer. When Form is selected, you must define the field that the Agent should fill in. Accuracy will be highest when your Prompt includes clear instructions on what data should go into each field.Form Prompts are very efficient because they allow you to group multiple outputs into a single Prompt. You can use Form prompts to extract many different pieces of information such as key details from a contract (e.g., parties involved, dates, amounts) by section.
Accuracy of Form Prompts will decrease with large (20+) numbers of fields. Recommendation: if the number of fields becomes too large, split your instruction into multiple Prompts.
Number Prompts should be used when the expected response is a single numeric value.
True/False (Boolean) Prompts should be used when the expected response is a true or false answer.
Classification Prompts are a special form of Text Prompt where the Agent is required to choose from a predefined set of categories. This is useful for tasks like sentiment analysis, topic categorization, or any scenario where the response must belong to a specific set of options.

​
Inputs to Your Prompts

In addition to the text instructions, you can provide additional information to your Prompts using file attachments or even the results of other Prompts.

​
Attaching Files

Attaching a file to your Prompt allows to provide longer context that may be impractical to include directly in the Prompt instructions. You can attach files of various types including PDFs, images, spreadsheets, and more. Some examples of useful attachments include:
  • CSVs of lookup tables or reference data
  • Guidelines or policy documents you want the Agent to follow
When attaching files, consider the following best practices:
  • Ensure the attached files are relevant to the Prompt and contain information that will help the Agent generate accurate responses.
  • If the file is large or complex, provide additional instructions in the Prompt to guide the Agent on which sections to focus on.
  • Use clear and concise language in your Prompt to specify how the Agent should utilize the information from the attached files.

​
Connecting Cloud Documents

If you wish to attach a document stored in a cloud service, you can do so by first adding an integration for that service (see Supported Integrations), then clicking that integration from the Prompt and selecting the file. Supported integrations include:
  • Google Drive
  • Notion (KMS)
  • SharePoint

​
Prompt Inputs (Chaining)

Sometimes a task is too complex to be handled reliably by a single Prompt. Kolena allows you to define and chain multiple Prompts together, where the output of one Prompt can serve as the input to another. This can be used in combination with other tools, for example:
  1. Use a Table Prompt to extract structured data from a document.
  2. Use a Data Analysis tool-enabled Prompt to perform calculations or analysis on the extracted data.
To use another Prompt as the input to your current Prompt, simply select it from the Prompt Inputs dropdown item or use the hotkey @ when editing the Prompt text.

​
Tools

Advanced tools can be enabled on your Prompts to enhance their capabilities. Each tool is designed to overcome a natural limitation of state of the art AI models by leveraging specialized techniques or services.

​
Data Analysis

Certain operations, such as mathematical calculations are best handled by specialized tools rather than relying solely on AI models’ pattern matching capabilities. Selecting Data Analysis for your Prompt will have the Agent write a piece of code which is then executed to produce the final output. This is especially powerful for tasks involving numerical data or complex calculations on tabular data. When You Should Use Data Analysis:
  • When you want to perform calculations, aggregations, or statistical analysis on data extracted from tabular data.
How Data Analysis Works:
  1. The Agent will analyze any input document(s) and your Prompt and write a piece of code.
  2. This code will be executed and the results returned as the final output of the Prompt.
When Data Analysis Is Not Suitable:
  • When you require natural language analysis. Recommendation: perform the task using two Prompts chained together, the first to extract the data and the second to analyze it using natural language.
If computation logic is complex, and you already have the code implemented in Python, you can provide the script as a Prompt attachment, and add instructions in your Prompt to guide the Agent to use it.For example, putting the desired processing logic in transaction_processing.py:
We can create a Prompt with the above code as a Prompt attachment, and automatically instruct the Agent:
Use the process function in transaction_processing.py, pass all transaction history in the document as history argument. Return the result as a table.
This can provide strong guidance for the Agent to run the exact computation.

​
Enhanced Vision

State of the art AI models excel in many vision tasks. However, for a specific subset of tasks, such as accurately reading checkbox states from documents, specialized vision models will provide superior accuracy. When You Should Use Enhanced Vision:
  • Prompts which require correctly reading checkbox states
How Enhanced Vision Works:
  1. The Agent will identify the regions of interest in the document or image based on your Prompt instructions.
  2. These regions are then processed using specialized vision models to extract the required information.
When Enhanced Vision Is Not Suitable:
  • When checkboxes exist on many pages for extraction (5+ pages)
Web Search allows your Agents to find information from the internet to answer questions with verified references. When You Should Use Web Search:
  • When your Prompt requires current information from the web or verification of facts.
  • When the information needed is likely to be found on publicly accessible web pages through a search engine.
How Web Search Works:
  1. Based on your Prompt, the Agent will decide if searching the internet is necessary.
  2. If so, the Agent will make search engine queries to gather information and refine searches for up to 10 of the most relevant web pages.
  3. The Agent will extract at most 120k words and analyze the content to answer your Prompt.
  4. The final response will use any documents plus these search results to generate a comprehensive answer with citations.
When Web Search Is Not Suitable:
  • Web Search cannot access private or paywalled content, nor can it download files like PDFs or images.
  • Web Search relies on search engine results and cannot access pages that block bots or crawlers.
Effective Web Search PromptsCraft your Prompts to steer the Agent’s search for information by:
  • Asking for current information (e.g., “What are the latest developments at Kolena in 2025?” instead of “What’s Kolena doing?”)
  • Using time-sensitive language like “latest,” “in the past decade,” “breaking news,” or specific time periods.
  • Requesting verification and fact-checking (e.g., “Verify the current stock price of Company-Name and cite your sources.”)
  • Hinting at specific sources or types of sources (e.g., “Search for information on Zillow about property prices in NYC” or “Find recent news articles about Topic-Name”)

​
Advanced Settings

​
Inline Citations

Embed file and page references directly within the output text, showing which sources were used for each part of the response. Use this when: You need to trace the source of specific information or when compliance requires detailed source attribution.
Inline Citations are only available for Text, Table, and Form output types. Inline Citations are enabled by default.

​
Ignore Run Files

By default, when a Prompt with Input Prompts executes, two sources of information are used as input:
  1. The files uploaded to the Run
  2. The results of any Input Prompts
When enabled, only the second source will be used as input for the Prompt. Use this when: You are chaining Prompts together in a multi-step workflow, and the initial Run files are no longer relevant to later steps. This helps ensure that later Prompts operate strictly on transformed or refined intermediate outputs, rather than being influenced by the original files.
Suppose your goal is to produce a final summarized report based purely on the results of earlier transformation Prompts rather than the original uploaded materials.Here is an example flow:
Prompt A: Extracts structured data from the uploaded files into a normalized format. Prompt B: Final summarization step.
If Prompt B still has access to the original Run files, the model may accidentally pull details directly from those files instead of relying solely on the refined output from Prompt A. To avoid this, you would enable Ignore Run Files on Prompt B. This ensures the final summary is based only on the transformed, cleaned, or reduced information produced by the previous Prompts.
This option is only available when Input Prompts are used.

​
Filter Documents (deprecated)

Deprecated: This feature is deprecated and may be removed in a future release. Avoid using it in new Agents.
The recommended way to filter documents is now a dedicated Document Filter step, which you define once and connect to any number of Prompts.
When enabled, the system automatically selects only the relevant documents for your Prompt. This can improve extraction quality and reduce noise when a Run includes many files. Use this when: There are many Run files, but only a few contain what you need. You want to exclude irrelevant files without manually selecting documents for every Run.
For best results, be explicit about which documents should be included (and, if helpful, which should be excluded). Mention the document type or identifying cues like filenames or keywords.Here is an example:
Only use invoices from 2026. Ignore receipts, statements, email threads, and invoices from other years.
This is only applied to Runs with more than 10 files or multiple files with more than 200 total pages.
This feature is unavailable when Ignore Run Files is enabled.

​
Control Steps

Beyond the Prompts themselves, you can add steps to your Agent that control what runs.

​
Document Filters

A Document Filter is a dedicated step you add to your Agent that narrows down which documents your Prompts will see. It is the recommended way to control document selection: you define a filter once, write instructions describing which documents to include, and connect it to the Prompts that should receive only those documents. Connecting a Document Filter to your Prompts optimizes both extraction quality and credit usage, since each connected Prompt only processes the documents that pass the filter rather than every file in the Run. When You Should Use a Document Filter:
  • When a Run includes many files but only some are relevant to a given Prompt.
  • When several Prompts should all operate on the same relevant subset of documents.
  • When you want to reduce credits by preventing Prompts from processing irrelevant files.
How Document Filters Work:
  1. Add a Document Filter to your Agent and write instructions describing which documents to include.
  2. In the filter’s Input field, choose what the filter reads: All Run Files (the default) or All Run Filenames.
  3. In the filter’s Output field, select the Prompts that should receive the filtered documents.
  4. When a Run executes, the filter analyzes the Run’s documents and passes only the matching subset to each connected Prompt.
  5. The filter’s result is shown as the number of documents that passed out of the total (for example, 3 / 12 passed).
Be explicit about which documents should be included (and, if helpful, which should be excluded). Mention the document type or identifying cues like filenames or keywords.Here is an example:
Only include invoices from the current year. Exclude receipts, statements, email threads, and invoices from other years.

​
All Run Filenames

By default a Document Filter’s Input is All Run Files, so it reads the contents of each file. Switch the Input to All Run Filenames to filter using only the filenames, without reading file contents. It is faster and uses fewer credits.
This works best when filenames are consistent and descriptive. For example:
Include files whose names contain Invoice. Exclude files containing Receipt or Statement.
If filenames are generic (scan1.pdf, document.pdf), keep the Input on All Run Files so the filter can read the contents.

​
Conditionals

A Conditional is a step you add to your Agent that decides which of your Prompts run for a given Run. It works much like a Classification Prompt. You define the set of conditions it can choose from (for example, Invoice, Receipt, Other) and write instructions describing how to choose between them. The difference is what the chosen condition does. When a Run executes, the Conditional picks the matching condition, and each Prompt you attached to that condition runs. Prompts attached to the other conditions are skipped. Like Document Filters, this improves both extraction quality and credit usage. Each Prompt only runs where it applies, and a skipped Prompt consumes no credits. When You Should Use a Conditional:
  • When your Runs contain mixed inputs, such as different document types, and different Prompts apply to each kind.
  • When a Prompt only applies in certain scenarios, such as when an optional section is present, or a review step that should only run when an earlier Prompt’s check fails.
How Conditionals Work:
  1. Add a Conditional to your Agent and write instructions describing how to choose the matching condition (for example, “Categorize the files by which company they belong to: Company A, Company B, or another company”).
  2. Define the Conditions the Conditional can choose from.
  3. Attach Prompts to conditions, either from the Conditional itself (each condition has an “If … then run …” row) or from a Prompt’s Condition field. Prompts left unattached always run.
Like any Prompt, a Conditional reads the Run’s documents, can use other Prompts’ results as inputs, and can be connected to a Document Filter. A Prompt can also carry conditions from more than one Conditional. These combine with OR, so the Prompt runs when any one of its conditions is met.
If a Run doesn’t clearly match any condition, the Conditional returns “No condition matched” and every attached Prompt is skipped. If you want certain Prompts to run in that case, add an explicit catch-all condition such as Other and attach them to it.
Avoid repeating the decision inside your Prompt instructions. Instead of writing “If this is an invoice, extract X, otherwise return nothing” into several Prompts, create one Conditional that decides once and attach those Prompts to its Invoice condition.
A skipped Prompt produces no output. If another Prompt uses a skipped Prompt as an input, it will see no value from it, so when a whole chain of Prompts belongs to one condition, attach every Prompt in the chain, not just the first.