Get Started

Introduction

Welcome to Inspire AI documentation portal. We welcome you as a beta user of this tool as we excitingly be working hard to improve the system and resolve bugs raised by you.

If you haven’t already, please fill in the invite form by clicking here. Doing so, will let us know that you are interested in testing our tool and we can create an account for you from the backend.

When you received your account credentials, head to ai.icubeutm.ca to start using Inspire AI. Please note that all of your data, credentials, chat history &… are being stored on a private server managed by ICUBEUTM. This means that your data will not be shared externally.

Key concepts

Text generation models

Text generation models, have been trained to understand natural and formal language. The inputs to these models are also referred to as “prompts”. Designing a prompt is essentially how you “program” a model, usually by providing instructions or some examples of how to successfully complete a task. Models can be used across a great variety of tasks including content or code generation, summarization, conversation, creative writing, and more.

Tokens

Text generation models process text in chunks called tokens. Tokens represent commonly occurring sequences of characters. For example, the string ” tokenization” is decomposed as ” token” and “ization”, while a short and common word like ” the” is represented as a single token. Note that in a sentence, the first token of each word typically starts with a space character. As a rough rule of thumb, 1 token is approximately 4 characters or 0.75 words for English text.

One limitation to keep in mind is that for a text generation model the prompt and the generated output combined must be no more than the model’s maximum context length. The maximum context lengths for each text generation and embeddings model can be found in the model section of the documentation.

Prompt engineering

This guide shares strategies and tactics for getting better results from large language models. The methods described here can sometimes be deployed in combination for greater effect. We encourage experimentation to find the methods that work best for you. In general, if you find that a model fails at a task and a more capable model is available, it’s often worth trying again with the more capable model.

Six strategies for getting better results

Write clear instructions

These models can’t read your mind. If outputs are too long, ask for brief replies. If outputs are too simple, ask for expert-level writing. If you dislike the format, demonstrate the format you’d like to see. The less the model has to guess at what you want, the more likely you’ll get it.

Tactics:

Include details in your query to get more relevant answers
Ask the model to adopt a persona
Use delimiters to clearly indicate distinct parts of the input
Specify the steps required to complete a task
Provide examples
Specify the desired length of the output

Provide reference text

Language models can confidently invent fake answers, especially when asked about esoteric topics or for citations and URLs. In the same way that a sheet of notes can help a student do better on a test, providing reference text to these models can help in answering with fewer fabrications.

Tactics:

Instruct the model to answer using a reference text
Instruct the model to answer with citations from a reference text

Split complex tasks into simpler subtasks

Just as it is good practice in software engineering to decompose a complex system into a set of modular components, the same is true of tasks submitted to a language model. Complex tasks tend to have higher error rates than simpler tasks. Furthermore, complex tasks can often be re-defined as a workflow of simpler tasks in which the outputs of earlier tasks are used to construct the inputs to later tasks.

Tactics:

Use intent classification to identify the most relevant instructions for a user query
For dialogue applications that require very long conversations, summarize or filter previous dialogue
Summarize long documents piecewise and construct a full summary recursively

Give the model time to “think”

If asked to multiply 17 by 28, you might not know it instantly, but can still work it out with time. Similarly, models make more reasoning errors when trying to answer right away, rather than taking time to work out an answer. Asking for a “chain of thought” before an answer can help the model reason its way toward correct answers more reliably.

Tactics:

Instruct the model to work out its own solution before rushing to a conclusion
Use inner monologue or a sequence of queries to hide the model’s reasoning process
Ask the model if it missed anything on previous passes

Use external tools

Compensate for the weaknesses of the model by feeding it the outputs of other tools. For example, a text retrieval system (sometimes called RAG or retrieval augmented generation) can tell the model about relevant documents. If a task can be done more reliably or efficiently by a tool rather than by a language model, offload it to get the best of both.

Test changes systematically

Improving performance is easier if you can measure it. In some cases a modification to a prompt will achieve better performance on a few isolated examples but lead to worse overall performance on a more representative set of examples. Therefore to be sure that a change is net positive to performance it may be necessary to define a comprehensive test suite (also known an as an “eval”).

Models

Inspire AI is powered by a diverse set of models with different capabilities and price points. You can also make customizations to our models for your specific use case by requesting a custom model via email to mohammad.tahvili@utoronto.ca

ChatGPT: Open AI’s ChatGPT 3.5 and 4.0 can understand as well as generate natural language or code. Additionally, these models are great for connecting to the internet and reviewing PDF documents.
Falcon: Strongest models that competes with ChatGPT, made by TII, that can understand as well as generate natural language or code.
Inspire AI: Custom model trained to help with business development & marketing. Created by ICUBEUTM, it supports ventures with their business strategy and development.
Llama 2: Open source model created by Meta that can understand as well as generate natural language.
Code Llama: Code Llama is an open source model created by Meta that can understand as well as generate code.
IDEFICS: AI model that can understand as well as generate natural language. Additionally, this model can understand images/media however will not generate any media.

Here is a table comparing the main models we offer to each other:

Feature/Model	Llama 2 70B	Falcon 180B	ChatGPT 3.5	ChatGPT 4.0	Inspire AI	IDEFICS
Parameters	70 billion	180 billion	~175 billion	~175 billion	~175 billion	114.9 billion
Max Context	1024 tokens	1024 tokens	2048 tokens	8192 tokens	2048 tokens	1024 tokens
Modalities	Text-based	Text-based	Text-based	Text-based	Text-based	Text-based & Multi-modal
Accuracy	High	Very High	High	Very High	High	High
Complexity	High	Very High	High	Very High	High	High
Speed	Medium	Medium	Fast	Fast	Slow	Medium
Efficiency	Efficient	Efficient	Efficient	Efficient	Efficient	Medium
Creativity	High	Very High	High	Very High	High	High
Cost	Free	Free	$0.002 / 1K tokens	$0.03 / 1K tokens	$0.002 / 1K tokens	Free

Feature list

Keep track of changes to Inspire AI. This feature list is maintained in a best effort fashion and may not reflect all upcoming changes being made.

Enhancement
Introduce an "upload PDF" button to help user experience.
Enhancement
When converting PDF to Text, the text needs to be invisible to the user.
Feature
Provide ability for the models to understand other file formats.
Feature
Provide models that can create/edit images.
Feature
Provide a model for different fields of expertise.
Feature
Provide a model to help in personal development growth.

User Guide

Conversation management

Start a conversation

To start a conversation, find and press the button that represents a pen:

This will open up a blank page where you can select your preferred model and start a conversation. Finally, find the input field typically labeled as “Ask anything” or similar. Enter your query or message and press ‘Enter’ or click the Send button.

Continue a conversation

Locating Previous Conversations: On the Inspire AI interface, look for a section usually on the left side. Or on the phone, by pressing the menu icon which is on the top left corner.

Accessing a Conversation: Click on the desired conversation from the list to view the previous interactions.

Deciding to continue a conversation vs. starting a new one

Continuing a Conversation: If your query is a follow-up or related to the previous conversation, it’s best to continue in the same thread/conversation for context.

Starting a New Conversation: If your new query is unrelated to the previous topics or if you prefer a fresh start, begin a new conversation. This can be done by either navigating back to the home screen or selecting an option to start a new chat, if available.

Delete a conversation

Selecting a Conversation: Navigate to the conversation you wish to delete. You can do this by placing your mouse/cursor on the conversation from the sidebar.

Deleting the Conversation: Look for an option like ‘Delete’, ‘Remove’, or a trash bin icon. Click on it and confirm the deletion.

Warning Deleting a conversation is an irreversible action, and the conversation history will be permanently removed from the server.

Change active model

Switching between models

Accessing Model Options: Locate an options menu, when starting a new conversation. Usually titled “Current Model”:

Selecting a Different Model: Click on the Current Model options menu. This will open up a box showing all the models available. Scroll through the list to find what you are interested in. Select the one you are interested. Make sure there is a visible checkmark next to the model you are interested in. Finally press the “Apply” button to apply the model to your new conversation.

Verifying the model you are using

Checking the Model Information: The current model in use is typically displayed at the bottom of the chat window. There you can find a sample text like below:

Model: Llama 2 · Generated content may be inaccurate or false.

Analyze image files

To analyze images, you need to insert the image in your prompt. At the moment we only have this feature active for one of the models called IDEFICS model.

Please first start a conversation and change the active model to IDEFICS. After changing the model, you will see a new button called “Upload Image” located at the top left corner of the chat window. By clicking on this button, you will be prompted to upload an image. After you have uploaded your desired images, please make sure to type in what you like the model to do with these images.

Analyze PDF files

To analyze PDF documents, you need to insert the document in your prompt. At the moment you can do so by dragging and dropping the document into the chat window of any conversation.

Note: It is recommended to use ChatGPT 3.5 for this as the large context window will allow you to upload large PDF documents.

Policies

Data security

Sharing conversation history with model creators

By default, the conversation history on Inspire AI is shared with the creators of each AI model used, except for ChatGPT models.
Users can opt out of this sharing by navigating to the “Settings” section and deactivating the option labeled “Share conversations with model authors.”
This feature is designed to improve model performance and user experience, while also contributing to ongoing AI research.

Conversation storage and control

All user conversations are stored on Inspire AI’s private server, which is managed and controlled by ICUBEUTM.
Users have the option to delete their conversation history. This can be done either by deleting each conversation manually or by using the settings page to delete all conversations in bulk.
For transparency and academic integrity, as well as for data training purposes, all conversations are stored unencrypted. This approach supports our commitment to advancing AI research and enhancing the capabilities of our AI models.
We take reasonable steps to ensure the security of your data on our servers. However, please be aware that no method of electronic storage or transmission over the internet is entirely secure.

Privacy

Personal information collection

Information you provide:

Account Information: When creating an account, we collect details such as your name, contact information, account credentials managed through Auth0, and transaction history.
User Content: Your inputs, file uploads, and feedback during service use are collected.
Communication Information: Any communication with us is recorded, including your name, contact details, and message contents.
Social Media Information: Interactions on our social media pages lead to collection of details like contact information.

Automatically collected information:

Technical Information: Includes Log Data (IP address, browser type), Usage Data (types of content viewed, actions taken), Device Information (device name, OS, identifiers), and Cookies.
Analytics: We use analytics tools to understand service usage, which involve cookie usage.

Use of personal information

To deliver and improve Inspire AI services.
For communication, research, and development of new features.
For fraud prevention, IT security, and legal compliance.
Business transfers and legal obligations may necessitate disclosure.
Aggregated or de-identified information is used for analysis and research.

Disclosure of personal information

Vendors and Service Providers: Shared with third parties for operational needs (e.g., hosting, IT services).
Legal Requirements: Shared when legally required or for protecting our rights and safety.
Affiliates: Shared within our controlled entities, under this Privacy Policy’s scope.
Business Account Administrators: If you’re part of an enterprise account, your information may be accessible to your account’s administrators.
User and Third-Party Sharing: Features may allow sharing of your information with other users or third parties.

User rights

Rights include accessing, deleting, rectifying, transferring, or restricting your personal information.
Withdraw consent for data processing at any time.
File complaints with data protection authorities.

Ethics

Ethical Considerations, Bias, Risks, and Limitations

Introduction to Ethical AI:
Inspire AI is committed to the responsible development and use of AI technologies. We recognize the transformative potential of AI and its impact on society. Therefore, ethical considerations are at the forefront of our AI model development and deployment.

Understanding Risks and Limitations:
Large language models, like those used in Inspire AI, are advanced yet carry inherent risks. Testing predominantly in English does not encompass all possible scenarios, which means that outputs can be unpredictable and may sometimes be inaccurate, biased, or objectionable.

Bias and Fairness:
Significant research, including works like Sheng et al. (2021) and Bender et al. (2021), has highlighted bias and fairness issues in language models. Inspire AI models, derived from these, can inadvertently produce content with stereotypes or inaccuracies. We are actively working to mitigate these issues through diverse data sets and inclusive model development teams.

Red-Teaming for Quality Assurance:
Our Red-Teaming efforts focus on identifying instances where models may generate incorrect, biased, or offensive responses. This rigorous process, involving diverse scenarios and continuous updates, aims to refine our models responsibly.

Bias Evaluation

Methodology:
Bias evaluation in Inspire AI is a multifaceted approach, primarily focusing on instruction-tuned model variants. Our methods include:

Red-Teaming: We simulate challenging scenarios to test the model’s outputs, considering the nuances introduced by text and image prompts.
Systematic Evaluation: Comparisons across gender and race axes are performed to identify and address any biases in model generations.

Intended use and scope

Intended for Research and Commercial Use:
Inspire AI models are designed for English language use in commercial and research environments.

Out-of-Scope Uses:

High-stakes decisions and scenarios with significant individual impact are not suitable for our models.
The model should not be used for critical automatic decisions, generating factual content, or any use requiring absolute accuracy.
Misuse, including spam, disinformation, harassment, or any form of abuse, is strictly prohibited.

Hardware and software ethics

Sustainable and Responsible AI:

Training involves custom libraries and cloud computing resources.
We are committed to offsetting our carbon footprint through comprehensive sustainability programs.
Ethical sourcing and energy efficiency are key considerations in our hardware and software choices.

User responsibility and transparency

User’s Role in Ethical AI:

Users are encouraged to use Inspire AI responsibly and stay informed about AI ethics.
We provide resources for continuous learning and understanding of AI’s societal impacts.

Transparency and Accountability:

Inspire AI is transparent about our models’ capabilities and continuously works to improve their accuracy and fairness.
We hold ourselves accountable for our AI’s impact and are committed to addressing any ethical breaches responsibly.

Citation

Citing responses from an AI model, can be done in various formats depending on the citation style you are using. Below are templates for some of the most common citation styles:

APA (American Psychological Association) Format

In-text citation:
(Insert Model Name Here, 2023)

Reference list entry:
Insert Model Name Here. (2023). Title of the input query or topic. Retrieved from [URL of the AI platform]

MLA (Modern Language Association) Format

In-text citation:
(Insert Model Name Here)

Works Cited entry:
“Insert Model Name Here.” Title of the input query or topic, 2023, [URL of the AI platform].

Chicago Manual of Style Format

In-text citation:
(Insert Model Name Here 2023)

Bibliography entry:
Insert Model Name Here. 2023. “Title of the input query or topic.” Accessed [date]. [URL of the AI platform].

Harvard Style

In-text citation:
(Insert Model Name Here 2023)

Reference list entry:
Insert Model Name Here, 2023. Title of the input query or topic. [Online] Available at: [URL of the AI platform] [Accessed Date].

——-

Note that the “Title of the input query or topic” should be replaced with a brief description of the content or the specific question asked. Also, if the AI model does not have a specific retrieval URL, you can omit the URL part and just mention the platform where the AI model operates.

Remember to always check with your specific citation guidelines, as there might be variations depending on the institution or publication.

Reporting

Server speed

If you have noticed that the platform is performing slower than usual, or if you are experiencing delays in receiving a response, please rest assured that we are working to address the issue. We currently use a 2GB memory, which may be a factor in the slower performance. However, we do have options to increase the memory and improve the overall functionality of the site. If you would like to inquire about increasing the memory or have any questions or concerns, please do not hesitate to reach out to us at the email address provided (mohammad.tahvili@utoronto.ca).

Report a bug

Do you see a bug that needs to be reported? Please fill in the form below so we can best understand the issue:

Request a feature

There are so many great ideas on where we can take this project. Feel free to share them with us by filling in the form below: