Developing apps based on OpenAI’s API may seem easy, but when it comes to assisting in reading documents, “accuracy” is the primary criterion for measuring the product’s usability.
Accuracy, is reflected in two aspects:
-
Accurate input: the input provided to ChatGPT needs to be accurate.
-
Reliable output: the answers should be traced back to the original text, avoiding ChatGPT’s fictional responses.
Accurate Input
To enable ChatGPT to understand the contents of a PDF, document intelligence techniques, such as document structure recognition and OCR, are used to convert the PDF to a sequence of paragraphs, tables, and images, which is then transferred to ChatGPT. However, part of the reason for inaccuracies and omissions in the content is due to errors that occur during the process of converting PDF to document structure.
At the same time, ChatGPT has a character limit on its input, so it is not possible to transmit all the content directly from the document. It is crucial to retrieve more relevant content to enable ChatGPT to answer the questions directly.
Recognizing and understanding tables is another challenge. The difficulties lie in table positioning, table structure recognition, and table understanding. Compared to other document reading assistants based on ChatGPT, ChatDOC is highly accurate in understanding tables.
After selecting a table or paragraph, the recognition result will appear in the chat box on the right. This is especially useful for financial statement analysis and other scenarios where it maximizes the reliability of AI answers.
Reliable Output
For AI’s responses, ChatDOC provides a tracing function: all cited information is given a source. Click on the page number, and you can trace back to the original tables / paragraphs, making fact-checking more convenient.
The authenticity of ChatGPT is widely criticized, and the tracing function in product design compensates for the shortcomings of LLMs.