DeepSights is designed to answer business questions grounded in trusted knowledge that is extracted from your proprietary documents, such as research reports and analyses.
To get most value out of DeepSights’ answers, it is important to keep its working principles in mind and to follow our guidance on how to formulate your questions.
DeepSights feeds on text from reports
To answer your questions, DeepSights interrogates the textual content that has been extracted from your corpus of research documents.
While reports often contain value information in data tables, charts, and infographics, these cannot be reliably recovered in the extraction process: arrangement of text, font sizes, colors, visual cues – all these might be important to correctly interpret the data and AI/ML is not able to do so yet1. Therefore, this kind of content is discarded and DeepSights is not able to provide answers grounded in tabular or chart data at this time2.
Consequently, DeepSights excels at retrieving findings from your knowledge corpus that are explicitly verbalized and described; it is generally less suited to extracting specific numbers from tables or charts.
DeepSights is not a chatbot
To ensure robust and reliable responses, DeepSights is built to follow a strict question/answer interaction. By requiring the input to DeepSights to be a direct question, we can automatically and reliably apply a wide range of contextualization and prompt optimization techniques when tasking the underlying AI to generate an answer. This enables us to elicit high-quality answers from the AI, irrespective of prompting proficiency by the end-user.
On the other hand, this means that the user cannot supply prompts, like with ChatGPT, but is restricted to posing direct questions3.
DeepSights provides adjacent information
When asking a specific question, you might find at times that DeepSights comes back with information that is related, but not fully answering to your question’s intent. For example, you might inquire about a brand’s perception of a specific target group and might receive a response that also includes statements about the general population’s perception.
This behavior is intentional and aims to surface supplemental and contextualizing information to give users a fuller picture of the relevant information found in the knowledge repository.
If in doubt, DeepSights won’t answer
Ensuring the trustworthiness of answers is DeepSights’ top priority. Therefore, if the evidence found is not clear and straightforward to interpret from the extracted text, DeepSights will refrain from giving an answer.
Based on experience across a variety of customers, ~30% of questions do not yield an answer. This is very much intentional to prevent AI hallucinations or misleading responses from being served.
Guidance for asking your questions
You might imagine DeepSights as a capable assistant who is generally proficient in market insights, but not intimately familiar with the details of your business (yet). Ask your questions like you would ask a human colleague who has recently joined the company.
Please review the guidance below to get optimum results from using DeepSights. Except for the first around the subject matter, these are not hard rules – disregarding them will still provide answers – but following them yields significantly better responses.
Focus on the right subject matter
DeepSights can provide answers from reports and is explicitly trained to respond only based on evidence it finds. Hence, DeepSights does not answer general questions like Wikipedia, nor does it provide company-internal directory information.
✅ What are the top trends in frozen food in the UK for the general population?
❌ Who can help me with a segmentation?
❌ Where are the market share dashboards?
❌ When is the next leap year?
Clearly state your intent
Providing a fully formulated question, as opposed to some keywords, helps clarify your intent and generate better answers for you.
✅ How do consumers perceive the Acme brand in terms of sustainability?
❌ sustainability Acme
✅ How are brand image and market share of Acme evolving in European markets?
❌ How is Acme doing?
Include relevant context
DeepSights is not (yet) familiar with all details of your business and your role. Help it to understand your context.
Avoid indirect specifications like “our” – they are understood but make finding the best source material more difficult.
✅ How is Acme performing in Italy?
❌ How are we doing in Italy?
Clarify jargon
Avoid company jargon or explicitly clarify it. DeepSights understands generally known concepts and their relations (such as that toothpaste belongs to the oral care category), but it is not familiar (yet)4 with your company’s specific lingo.
✅ How is the segment “type B” responding to the top-tier offers (TTO)?
❌ How does type B respond to TTO?
Avoid instructions or chatter
DeepSights has an elaborate internal structure to generate AI responses that are tailored to the question / answer use case. Attempting have a casual chat, to provide instructions on style, length, output format, or other task guidance will not yield desired results, but rather degrade answer quality.
✅ What are top trends in frozen food in the UK for general population?
❌ … in three paragraphs?
❌ Write an email outline about …
❌ List all reports that talk about …
❌ How are you?