SquirroGPT#
This page provides an overview of SquirroGPT and some tips on how to use it.
The information on this page is directed towards end users. If you are a project creator seeking to set up a new SquirroGPT project, see the SquirroGPT Data Guide or the SquirroGPT Web Guide.
What is SquirroGPT?#
SquirroGPT is a retrieval-augmented generation (RAG) AI application powered by an underlying large language model (LLM). It is the first enterprise-ready generative AI application geared towards organizations searching for a natural language tool to conversationally interact with their data while meeting data security and privacy requirements.
Put in simpler terms, SquirroGPT is a digital assistant that uses cutting-edge AI technology to allow you to securely and privately chat with your own organizational data or allow users to securely and privately chat with your website content through an embedded digital assistant.
Introduction#
SquirroGPT allows you to chat with your own data, precisely and securely. While it operates similarly to other popular digital assistants, it’s important to understand the differences between SquirroGPT and generalized chatbots like ChatGPT, Bard, Claude, and others.
SquirroGPT is not designed to answer general knowledge questions or apply logic-based reasoning to user queries.
The marketplace is already flooded with generically-trained chatbots that perform these tasks.
Instead, SquirroGPT allows you to connect your own data, including Enterprise data or your own website data, and conversationally interact with it, privately and securely.
SquirroGPT also provides sources for its answers, allowing you to quickly understand the context of its answers by viewing a wide range of document types directly within the application.
Chat With Data and Chat With Web#
SquirroGPT is offered as two separate subscription options: Chat With Data and Chat With Web.
Chat With Data allows you to connect your own enterprise data to SquirroGPT using built-in Squirro connectors like the Microsoft SharePoint or Dropbox connectors.
Chat With Web allows you to index your website contents in SquirroGPT and embded a digital assistant on your site that allows website visitors to conversationally interact with your website and be provided with answers and sources.
Reference: To learn more about these plans, see SquirroGPT Pricing.
Take SquirroGPT for a Test Drive#
If you want to quickly see what SquirroGPT can do, you can now Test Drive a SquirroGPT application.
This allows you to safely and securely upload your own data or crawl your own website and see how SquirroGPT works.
To learn more, see How to Launch a Test Drive.
Tips for Maximizing Your SquirroGPT Experience#
SquirroGPT is a powerful tool that can be used in a variety of ways. Here are some tips to help you get the most out of your SquirroGPT experience:
Use Document Wording Where Possible#
The closer you word your prompts to how your connected documents are worded, the better your results will be.
If you know your documents talk about stock earnings in terms of “quarters”, use the term “earnings per share”, and generally use company stock symbols instead of names, here are a few recommended and not recommended prompts as examples:
What was Tesla’s stock earnings?
What was Tesla’s earnings in the last period of last year?
What was TSLA earnings per share in the fourth quarter of 2022?
The More Specific the Prompt, The Better#
The best prompts are detailed and specific to content that can be found (or suspected to be found) within the project’s connected documents.
Being detailed and specific also increases the likelihood of SquirroGPT returning quality sources.
If you are wondering if Squirro has any professional certifications, for example, it is better to be specific in your wording, as the following chat demonstrates:
As can be seen in the chats above, originally SquirroGPT was unsure about the type of certifications being asked about.
However, once given a specific “ISO” query, it was able to provide the correct answer along with accurate supporting sources.
Prompt Suggestions#
The following is a list of prompt techniques that can be used to improve the quality of SquirroGPT responses.
You can ask SquirroGPT to explain answers in ways that appeal to different audiences and levels of precision.
This includes the following:
Explain in detail
Explain like I’m 5
Explain with examples
You can ask SquirroGPT to provide answers in a specific style. This is an alternative to the Explain technique.
Try the following:
Answer in formal style
Answer in informal style
Answer in descriptive style
You can ask SquirroGPT to provide answers in a specific format, though not all formats supported by ChatGPT itself are supported by SquirroGPT (e.g. charts, PDF).
Try the following:
Format the answer as a list
Format the output as HTML
Format the answer as a table
The example below shows how being specific with prompts can return a table in the exact format you specify:
Features Not Officially Supported#
Although some of these features may work sometimes, they are not officially offered or developed as of Q3 2023.
The fact that they may partially work can create the misconception that they are supported features that are buggy, which is not the case.
Summarization#
While ChatGPT itself can summarize blocks of text, this is not a feature currently available on SquirroGPT.
The technical reason relates to the token-based way SquirroGPT queries the underlying LLM.
Logic-Based Operations (e.g. Analyses, Predictions)#
SquirroGPT is designed to retrieve answers, not perform logic-based operations on a set of documents.
For example, if you ask SquirroGPT to predict the future price of a stock, it will not be able to do so.
However, if you ask SquirroGPT to provide you with a given stock price as of the end of 2022, it will be able to do so (assuming that data is contained within your connected documents).
Transformative Outputs#
While SquirroGPT can provide basic formatting when it answers, it is not designed to transform documents into other formats or for other use cases, such as “Convert this document into a set of RFI responses”.
Multi-language Support and Translations#
While SquirroGPT may provide good translations or accurately answer in non-English languages, English is the only officially supported language.
Limitations#
SquirroGPT is a powerful tool, but like all generative technology powered by LLMs, it has its limitations, including the following:
Sources#
Although Squirro asks the LLM to indicate which snippet of provided information it used to formulate an answer, the LLM does not always respond, and when it does, is not always accurate.
These issues are out of Squirro’s hands to fully correct and represent known limitations of LLM-powered applications. There is no way to guarantee that sources are returned by the LLM, or that it will be 100% accurate when it is returned.
If you are trying to encourage SquirroGPT to return sources, be specific and use document wording wherever possible.
Highlighting#
For the same reasons as listed above (the LLM does not always tell Squirro which snippet was used, and is sometimes wrong), highlighting will not always be 100% accurate.
Additionally, there is a known issue with the PDF viewer that prevents highlighting being applied in multiple areas of documents.
Hallucinations#
SquirroGPT strives for accuracy, but sometimes hallucinations can occur due to the nature of the technology. Confirm answers using the provided sources. If you find a mistake, please let us know to help refine SquirroGPT and produce more accurate responses in future.
Complex Charts and Visuals#
SquirroGPT should not be relied on to extract information from complex charts and visuals such as infographics.
Lengthy Prompts#
If there is information SquirroGPT needs to query, that text should be uploaded as a document, not pasted into the chat.
Code Snippets#
Code snippets are not always produced accurately. The more bespoke the code, the larger the probability of issues.
How It Works#
The following diagram shows how user chat queries are converted to answers via the interaction between SquirroGPT and the underlying LLM that generates responses based upon provided source candidates.
Note
This shows the general logic of the interaction between SquirroGPT and the LLM, it is not a technically precise process-flow diagram.
Squirro Terminology#
Chat Queries: Term that refers to the inputs SquirroGPT users submit. Also known as user prompts, user questions, or user queries.
Source Candidates: List of documents that Squirro retrieves using a combination of keyword and vector search that serve as potential reference sources. Up to 10 candidates are retrieved.
Sources: The documents that answers were extracted from. The LLM indicates to Squirro which documents were used, then Squirro attempts to highlight relevant passages related to the answer.
LLM Prompt: Data that Squirro passes to the LLM, which includes the chat query and instructions, which the LLM uses to generate the response that is eventually provided to the user.
Reference: For more Squirro-specific terminology, see the Squirro Glossary.
Industry Terminology#
Embeddings: Refers to a technique of representing words or tokens as vectors in a high-dimensional space. The idea behind embeddings is to convert categorical, symbolic data (such as words) into numerical data that the model can process and learn from. Embeddings help models capture and understand the nuances of language, including things like word meaning, context, and even grammatical roles.
Hallucinations: Refers to a situation where the model generates information or outputs that are not based on the data it was trained on or provided. These are typically facts or assertions that seem plausible but are not accurate or reliable.
LLM: A Large Language Model (LLM) is an AI program based on transformer architecture, specifically designed to process and generate natural language. It utilizes machine learning principles, specifically deep learning, and is trained on large amounts of text data.
RAG: Retrieval-augmented generation is a type of natural language generation that combines the use of a retrieval model and a generative model. The retrieval model is used to retrieve relevant information from a database or a set of documents, and the generative model is used to generate a response based on the retrieved information.
Tokenization: Refers to the process of breaking down text data into smaller pieces, referred to as “tokens.” These tokens usually represent words or characters, depending on the type of tokenization used. For example, if the previous tokens correspond to the sentence “The quick brown fox jumps over the”, the model might predict that the next token should be “lazy”, because “The quick brown fox jumps over the lazy dog” is a common sentence.
Reference: For more industry terminology, see the Glossary of Industry Terms.
Contact Squirro#
Looking to get started with a SquirroGPT trial? Try it for free by visiting the SquirroGPT Free Trial Page.
If you’re looking for technical assistance, don’t hesitate to reach out to the friendly folks at Squirro Support.