Chatgpt Parameters

Changes from default values are indicated with color cues.

• max_token

  The maximum number of tokens that the AI model is allowed to generate in response to a particular prompt. May affect the cost per API call and the time it takes to get a response.
  model dependent

Related Token limits

• temperature

  Governs the randomness of responses. Higher values create diverse but potentially less focused replies, while lower values yield more focused and predictable answers.

  0 to 2

• Controls model's response predictability
• Lower temperature = more common responses
• Higher temperature = less predictable, more creative responses
• Common parameter in most LLMs

If we revisit the "The apple is..." example, with a lower temperature, the model is more likely to predict a common descriptor like "red". However, when you increase the temperature, it amplifies the chance of less common descriptors being chosen, such as "ripe" or "delicious".

• top_p

  Controls the word range for the AI's response. Lower values make it stick to more common words, whereas higher values allow it to consider a broader vocabulary.

  0 to 1

• Works with nucleus sampling to control response predictability
• Lower top_p = exact, fact-based responses
• Higher top_p = more diverse responses
• Common parameter in most LLMs

Note Generally recommended altering this or top_p but not both.

• Frequency Penalty

  This setting discourages the model from using commonly occurring words and phrases in its response, regardless of the specific context.

  -2.0 to 2.0

• Discourages common words/phrases
• High frequency_penalty = creative/unique, but possibly less straightforward responses

The frequency penalty discourages the model from using words or phrases that it understands to be common or frequently occurring. This understanding is based on the vast amount of text data from the internet that the model was trained on, not just the immediate context of the conversation.

By increasing the frequency_penalty, you're asking the model to generate responses that avoid these common words and phrases, leading to outputs that might be more creative or unique. However, a high frequency penalty can also make the model's responses less straightforward or harder to understand.

• Presence Penalty

  This setting influences how much the model should stick to or diverge from the words and phrases present in the current input or context.

  0 to 1

• Relates to current input/context
• High presence_penalty = encourages use of input-specific words/phrases, helps model stay on topic
• Low presence_penalty = introduces new words/ideas/topics, resulting in more varied responses, but possibly less related to initial input

In contrast to frequency_penalty, the presence_penalty directly relates to the current input or context. When set high, it encourages the model to use or stick to the words and phrases that have been mentioned in the conversation or input, helping the model to stay on topic.

Note Don't worry about remembering the values on Únacog AI, we have simply these values with a slider bar. For more information on these parameters, visit OpenAI's documentation.

Users link

The user section shows everyone who has access to the document.

The presence indicator displays the user's current status. A green dot means they currently have a browser open with the session. A gray dot means the user is no longer viewing session (that their brower is no longer connected).

The yellow and orange dot indicate that the user has submitted a prompt and is awaiting an AI response. Yellow is for any prompt submission, while organge indicates a parameter other than Max_tokens or model was changed from default.

Give access or share a session by simply giving someone a link

Use the "link" icon to copy the session URL to your clipboard.

If you notice the user section lighting up with a color anything other than white, this should cue you that this is a shared session.

Recent

Get quick access to previous sessions

The recent section provides access your last six sessions. Navigate to your dashboard if you would like to view or manage other previous sessions

Document Options link

You can access document options multiple ways: document creation, session, and within the dashboard for each individual session.

Document Options enables users to exercise a high degree of control over their sessions and has three tabs

Aside from the export tab, most of document options are available to users initially when they create the session.

Export Tab

This tab focuses on export options for your sessions.

You can choose to export the entire chat history or select specific parts to download, providing flexibility in managing and preserving your session data.

You can export your chat history in the following formats Text, HTML, JSON, CSV.

Note CVS and JSON can be uploaded to clone a session

Options Tab

UNDER CONSTRUCTION

This is where the majority of the session control tools reside

Send email

• Edit title: Allows you to modify the title of your session.
• Edit system message: Modify the system messages for a personalized experience. A system message gives context to a session without invoking a response.
• upload: Use this feature to upload and integrate previous prompts into your current session. Make sure to use JSON or CSV.
• Delete Session: Remove a session permanently.
• Archive Session: If you want to preserve a session without any further prompt requests, use this option to make the session read-only.
• Clone Session: Create an identical copy of the current session. This is especially useful when you want to try different approaches without modifying the original session.
• Token Usage Stats: Gives you a detailed report on token usage for the session, helping you to monitor and manage your resource usage more effectively.
• Clone Packet: Click to get a JSON of the packet info sent to API

Threshold Dialog

The Threshold Dialog is a valuable feature located within the Document Options. This handy tool provides crucial data about your session, particularly concerning maximum token usage. It's designed to offer you a clear and comprehensive view of your token count, so you have better control over your interaction with the AI.

• Max Token

  This displays the parameter setting set in the Engine, which governs the maximum token count for an AI response.
• New Prompt

  This shows the token count of the current input in the text area, which hasn't been submitted yet.
• System Message

  This reveals the token allocation used in the system message.
• Prompt Count

  This displays all prompts as a ratio between selected and unselected responses, giving you a snapshot of your prompt selections.
• Request Size

  This reflects the current token count in the context history (i.e., selected prompts), plus the value set for max tokens.
• Model Limit:

  This indicates the token limit for the selected model.

In addition, the Threshold Dialog provides an option to automatically exclude prompts that exceed the API limit, a function that is often performed invisibly in other app interfaces. Remember, understanding how tokens are used and managed in the interface is key to maximizing your interaction with the AI model.

Related How API max token limits work

With the Threshold Dialog, we strive to give you the insights and controls needed to truly shape your AI interaction. Happy experimenting!

Owner Tab

As an owner, you have additional tools at your disposal:

Usage limits per session

• Add Labels: This feature allows you to classify your sessions using custom labels, which aids in the organization and filtering of your documents.
• Owner's Note: A space for you to write notes related to the session. These can be reminders, observations, or any information pertinent to the session.

Profile Setting link

Overview

The Profile Options, accessible from both the chat interface and session dashboard, is an essential tool for personalizing and controlling your user experience in Únacog. The dialog presents the user's registered email at the top and the sign-out option at the bottom. It's structured into three tabs: User, Labels, and Account.

User

Use this tab to personalize your profile by changing your name and profile image

• Name: Use the pencil icon to manually change your name, or the dice icon to generate a random one.
• Profile Image: Upload a custom image or choose to generate a random mascot image.

In addition to these, there are four checkboxes for customizing display settings:

• Large Text: Check this box to increase the size of the chat interface text.
• Monospace Font: Switch the font family to monospace for chat prompts and responses.
• Auto Exclude: This feature automatically excludes the oldest prompt in your chat session to avoid going over the API token limit for each ChatGPT model. This option is also available on a session basis inside document options > threhold

Labels

This section presents default labels (Personal, Business, and Archived) that you can assign to sessions during creation or afterwards. You can modify these labels according to your needs for better organization and filtering of your documents.

Account

This tab provides critical information about your usage and subscription.

Token Usage History: Provides a detailed breakdown of token usage by day, month, year, and all time based on sent messages and replies from the AI.

Monthly OpenAI Token Usage: Displays the total token limit and your usage for the month.

New Email: Allows you to migrate your document data to another email account.

Subscription Level: Shows your current subscription level. During the beta phase, all users are identified as Pre Release

Foundations link

Getting to Know Large Language Models (LLMs)

Ever wondered how a Large Language Model, like GPT-3, comes up with its responses? Well, let's pull back the curtain a bit.

Think of LLMs as super-smart word predictors. They use the words you've given them to guess what word should come next. For instance, if you say, "The apple is...", it might guess "red" because that's a typical subsequent descriptor in sentences it's encountered during training.

But here's the thrilling part - it's not always "red". These models apply something called a probabilistic approach, which implies the same phrase could yield different responses. So, sometimes, the apple could be "ripe" or even "delicious" - it all hinges on the context!

Understanding how these models work helps us get the best out of them, whether we're teaching a class, exploring a research question, or just having fun. And for those of you who love diving into the nitty-gritty, understanding this foundational concept is key when it comes to fine-tuning how the AI responds - do you want it to be more creative or less repetitive? You'll get the hang of it!

Understanding and Managing Token Limits

Each Large Language Models (LLMs) has a token limit. GPT-3.5 Turbo's limit is 4096 tokens per conversation. This limit dictates the AI's context, meaning once reached, the earliest parts of the conversation are omitted from the AI's consideration.

In many AI interfaces, this isn't made explicit, potentially leading to confusion if the AI responses lack context from earlier messages. Únacog AI addresses this by keeping users informed about their token usage in a session, ensuring they are aware of the AI's current context.

Únacog will alert you to exclude previous messages when we detect you may go over the model's token API limit

Understanding and managing token limits can significantly enhance your AI interaction. You should become aware of how many tokens are used in each message, and what the limit is for the model you are using.

One strategy for managing content is to select which prompts are included or excluded from the AI's context. This gives you control over the information the AI considers, optimizing your interaction within the token limit, and improving response quality.

Understanding token limits will also help you gauge how much it will cost to run something. Usage can add up quickly on more complex models: chat-gpt-3.5-16k is about 2x the cost of gpt-3.5-turbo. While gpt-4 is about 10x more expensive than gpt-3.5-turbo

Tokens management is a important concept to grasp for advance users, learn more by reading more on Tokens in this OpenAI Article

FAQ link

Coming Soon!