Converse

Converse allows you to have a conversation with your documents, getting on-demand information from documents of nearly any type or format.

Files you upload are stored to Instabase Drive by default. Instabase Drive includes up to one terabyte of storage and is included in your AI Hub account. Optionally, you can connect external storage options. You can upload the following file and batch sizes to Converse:

• Files up to 50 MB or 800 pages.

• Up to 100 MB per upload batch.

• Up to 100 documents per conversation.

For more information, see AI Hub’s upload and storage limitations.

Conversing with your documents

Ask any question of your documents, extract data from tables, lists, and checkboxes, and get your results in a variety of formats, including tables, charts, and graphs. Converse can identify information from single or multiple sources and select the relevant information for the response. You can query up to 50 documents at one time to find information or compare data across documents.

For example, for a single document, you might ask:

• What were the total tax deductions from this paystub?

• What was the highest earning fund in this 401k statement?

Or upload multiple documents and ask:

• Does this invoice rate match what’s stated in the contract?

• What conditions need to be met for this claim to be approved based on the guidelines?

• Do all these documents have the same address for James Smith?

When you start a conversation, it includes only the first document you’ve uploaded to that conversation. Your conversation scope — the document set that you are conversing with — is displayed under the chat box. You can ask about any or all the documents in the conversation scope.

You can always adjust your conversation scope between prompts:

• To switch your conversation to a different single document, select the document from your conversation’s document list.

• To add a document to the conversation scope, select the document from the conversation’s document list and click the + icon.

• To remove a document from the scope, click the x on the document name in the scope bar.

• To clear the scope, click the + icon in the scope bar and click Clear scope in the dialog.

For conversations with multiple documents, Converse uses the multistep model by default, as it returns more accurate, robust results on multiple documents. For details about each model’s strengths and capabilities, see Choosing a model. Multiple document conversations are supported only with the multistep or advanced models. Multiple document conversations also have a few other limitations.

Extracting objects

You can extract information from tables, checkboxes, and signatures in documents using specific settings and prompts. Converse can extract information for a single specific table or for all tables in the document.

For best results, enable object extraction in the conversation settings. Table extraction in Converse is a public preview feature and has some limitations.

When you enable object extraction, any documents already in the conversation are redigitized. Redigitization can take some time, depending on the number and complexity of files in the conversation.

Conversation settings aren’t yet available on the mobile version of Converse.

1. Mouse over the My Conversations sidebar and select the conversation for which you want to change settings.

2. Click the settings (gear) icon.

3. Enable Object detection: Multipage tables and checkboxes.

Extract tables

To extract information from all tables in a document, begin your prompt with extract all tables. To get results from multiple tables in a specific format, specify either Markdown or JSON. For example:

• Extract all tables and return in Markdown

• Extract all tables and return in JSON format

To extract information from a single specific table in a document, include the title or header of the table in your prompt. For example:

• Extract the transactions table for the month of January 2023

• Extract the monthly transaction summary for the month of January 2023

Converse returns single-table extraction results as a table, which you can copy or download as a CSV file.

You can also filter columns or rows, sort columns, and perform other manipulation of table data. For example:

• Extract transactions and filter for amounts greater than $1,000

• Extract transactions and return results for 01 May through 15 May

• Extract transactions table and sort amounts from smallest to largest

• Extract transactions and add a column Flagged with values set to Yes if the debit is greater than $70

Highlight tables

When you enable object detection, you can highlight all detected tables in the document. Highlighting tables enables you to view an enlarged version of a given table, copy the table in JSON format, or download the table in CSV format.

To highlight detected tables, click the highlight icon at the top of the document viewer. To view, copy, or download a highlighted table, click the grid icon on the table and select the action.

Extract checkboxes

You can extract information from checkboxes in single or multipage documents.

• For a group of checkboxes with a label, such as the Filing Status field on a tax form, use a prompt asking about which checkboxes are selected, such as What filing status is claimed?.

• For a standalone checkbox, use a prompt that indicates whether the checkbox is ticked. For example, Is the filer claiming capital gains or losses?

Extract signatures

You can extract information about signatures, including whether a document is signed, who the signer was, and the signature date. Extraction of signature images is not available.

For example, you can ask Converse:

• Extract all signatures

• Is this document signed?

• Who signed this document?

• Are these documents signed by the same people?

Result formats

Converse can return responses in plain text or in other formats, including rich graph formats such tables, lists, charts, and code blocks. These result formats can be copied or converted into other formats and downloaded.

Available formats for Converse responses include:

• Tables, also downloadable as CSV files.

• Code blocks: More than 25 formats available, including JSON, Python, bash, and JavaScript.

• Charts: Line, bar, column, pie, scatter, and multi-axis. Charts are also downloadable as CSV, PNG, or SVG files.

To get results in a specific format, ask for that format in your prompt to Converse. For example:

• Identify all beneficiaries in this trust document along with the beneficiary type, and return the results as a table.

• Show the payroll deductions as a pie chart.

• Show the percent increase in net sales from 2019 to 2022 as a bar chart.

Not all result formats are available in all models. To see available copy and download options for a given response, mouse over the corner of the response.

How to ask great questions

Converse gives better answers when you ask better questions. These tips can help you to get the most out of Converse.

Provide clear instructions

Use verbs like extract, identify, calculate, find, explain, and summarize, depending on your documents and question. If you want the answer in a structured format such as table or in JSON, specify the result format explicitly.

Ask the model to think step-by-step

Especially for complex tasks, ask Converse to think step-by-step by adding “explain step by step” to the end of your prompt. When the model focuses on each task individually, it improves the accuracy of each response.

Provide additional context

Give Converse more context about your question to help it better understand what information you need.

Use more advanced models for complex documents and queries

For long documents or questions that require complex reasoning, try the advanced or multistep models.

Choosing a model

Converse offers default, advanced, and multistep models. For conversations with a single document, the default model is used. For conversations with multiple documents, the multistep model is used. You can change the model in the chat box.

For details about model capabilities, see Choosing a model.

Limitations to keep in mind

Keep these limitations in mind as you work with Converse.

Chat history

For now, Converse has no memory of the questions that you’ve already asked. It’s crucial that your question includes any relevant information from previous questions. For example, questions like “Give a shorter answer than your last one” don’t work. Ask the question again with more explicit instruction, such as, “Who wrote this document, answer succinctly.”

Long documents

Converse can process and answer questions about very long documents. To do this, Converse breaks down the document in chunks and uses the most relevant chunks to answer your question. Sometimes, Converse might select the wrong chunks, or the data you’re requesting might be spread over more chunks than Converse considers. Be mindful of this — the page numbers in Converse’s answers tell you which pages Converse is using to answer your questions.

Hallucination

Sometimes, AI models “hallucinate” details that aren’t present in the original context or content. This usually happens when the model is asked to infer or speculate. To avoid this, ask clear and specific questions. Use your own knowledge of the document and check the page numbers Converse cites to double-check Converse’s answers.

Biases

The AI models used by Converse are trained on a large dataset, and therefore might unintentionally reflect biases present in this data. Bias is more likely to show when the model is asked to infer or speculate. When interpreting responses, be aware of potential biases, especially those that pertain to sensitive subjects.

Mathematics

Large language models aren’t good at complex mathematical tasks. Keep this in mind if you try solving complex algebra, geometry or statistical analysis problems using Converse.