The Lite Mode page is the Pragatix interface for the end user.language

It contains the list of chats (Assistants) the admin configures for the user based on group membership.

The Lite Mode has company chats connected to your company data source and a public chat to ask questions that do not need company data sources.

Overview

LiteMode has the following types of chat services:

General chat

In this chat, by default, you get answers from public (pre-trained) knowledge. Use this if you know the answer does not need company data, like generating code, planning a trip or improving text which is part of the prompt.
It will get better performance as it will skip trying to find the answer in the company’s knowledge.

You can also upload a file or files to this chat and get answers from those files. Once you upload files, the chat will give priority to answers based on those files. Only if they do not have the answer will it give an answer based on Public knowledge and indicate this was not from the files.
Press the company knowledge button to use the General chat and get answers from all of the files you have access to in the company.

Company and Public knowledge settings

When uploading a file or selecting company knowledge, you can choose whether to include public knowledge in the answer.

Public answer when asking on Public and Company:

If both company and public knowledge are selected, the chat will use both public or pre-trained information and company sources.

When uploading a file, the company label changes to “Selected content”

or selecting content, the system will respond based on the selected content and will add "public" if selected.

The company button means that the system will search all company knowledge.

No answer when asking on company data only:
If you choose not to use public knowledge and rely only on a file or company data, the system will not provide an answer unless it can confirm the source is from the company. In that case, you will see the following message:

No answer when asking on pubic or private and public

When the AI does not return any answer even though public is included , you would see the following message. The AI is instructed to answer accurately. In this case it decided not to provide an answer, as it didn’t have enough relevant knowledge to provide an accurate answer.

• Assistant Chats – Chats configured by the admin for specific company content,
designed to be experts in that area.

• Assistant Analysis Agent- Chat that is configured by the admin to analysis a DB

• My Analsys Agent - An AI agent that can be used to analyse Excel and some other funtions such as search the internet or modify files

• Smart Search - Search through all company content to find exact and similar results.

Uploading files to general chat

Duplicate/ existing documents

When you upload a file to the general chat, the system searches for an identical document that has already been uploaded. When such a case is identified, you will see a page showing a list of documents, with info on who uploaded them and when and comments if added.

Language support settings

The are 3 levels of language settings

1. UI

2. Chat language

3. Content language

UI language and chat response

These are set by default by the admin and can be changed by the user:

Determining Response Language

Chat Primary Language - defines the language that chat responses will be used in.

The admin sets the default value in the account settings, and the user can change it in their workspace settings.

The system automatically detects the language of each query: if the query is identified as English, the response will be in English; otherwise, the response will be in the primary language of the chat.

This ensures responses align with the user’s input while maintaining a consistent default language.

Language Examples:

If the Primary Chat Language is set to Spanish,

• Query: “¿Cómo configuro mi cuenta?” → Response in Spanish

• Query: “How do I configure my account?” → Response in English

If the Primary Chat Language is set to German,

• Query: “Wie ändere ich meine Einstellungen?” → Response in German

• Query: “What are workspace settings?” → Response in English

Content primary language

In addition to the chat language, the admin can set the primary languages of the content. This helps keyword search to be done in all language sets regardless of the UI and chat language

Show users interested in a document

When the system lists source documents, it shows the users who also showed interest in this document. This is based on any prompt we did directly in that document and on using that document to answer a prompt. Clicking the user's image opens a list sorted by the number of times that user has used it- Top used by and the user who has created/uploaded the file

Instructions

Response Instructions

The end-user can customize how the chatbot responds by selecting a pre-built instruction set. Use the Choose Style dropdown to pick a response style that best fits your needs. Options include:

• Competitor Analysis – Get structured insights and comparisons.

• Creative Writing – Elevate storytelling and imaginative responses.

• Customer Support – Provide clear, helpful, and empathetic replies.

• And More – Explore different styles to match your specific requirements.

Using Prompt Startups

The end-user can quickly ask questions using Prompt Startups, which are located below the chat

Public chat

You can ask questions based on the AI's pre-trained knowledge, including improving English, writing code, and planning trips.You can also upload a file from your computer and ask questions about it.

Company chat

Get answers from all company data you have access to.

At the bottom of the chat, the files on which the chatbot based its response are displayed. You can click on them and review them.

Company Assistant Chat

Under this tab, you can view all the chats you have access to. Each chat focuses on a unique topic and is based on different sources of information.

Smart Search

Use this feature to find content and documents within your company that match your search terms.

While the regular chat provides direct answers, this search returns a list of the most relevant items based on your query. It uses keyword matching and AI-powered similarity analysis to deliver the best results.

The results are sorted based on the file's relevance to the requested topic. You can view the files by clicking on them. Additionally, you can see the file's source and the specific text that determined its relevance.

Use the icons on the right to perform the following actions.

• Preview: View the file with AI-highlighted exact and similar words.

• Start Chat: Ask questions about the file directly from the results.

• More Info: See details like Title, Type, Upload Date, and Uploader.

Lite Mode FAQ

Public chat sources

If you ask a question without including a file - the source of the answer will be public knowledge
If you include a document - the source will be limited to that document only
From version 2.10.1 (April 25) you should use the company chat to upload files. In such case the chat will answer only on that file.

Can I upload a file into the company chat?

When you upload a file to the company chat you will get only answers on that file.

Is it possible for users to get different answers to the same question?

Yes. The system generates responses based on the content each user can access within the company's general chat and assistants. For example, one user might have access to a SharePoint folder with relevant information that another user does not, resulting in different answers. The AI’s creativity setting (temperature) may also introduce variations in its responses. In future versions, you can adjust this creativity parameter yourself.

Deep and quick modes

Balancing speed and quality of responces you can chose to the chat to user deep more or quick.

When selecting Deep, the system attempts to send more content and get a longer response.

When uploading a file or files - all of them are sent always because the system always uses deep in this case.

When asking on all company data or a collection, you can decide whether to use deep or quick.

When selecting quick - the system searches for the most relevant chunks (using Stuff algorithm)

Advanced options

Under Advanced option in the General chat, you can find the following options:

1. Prompt mode:

   1. Native mode- chosing this means no system prompts are sent to the model and the prompt with the files are sent as is.

   2. Minimal Prompt Mode- In this mode, most of the product system prompts are removed. Only instructions sent are to answer the question using only the provided documents, consider the conversation history, and clarify if the answer comes from the documents or my general knowledge. Recommended to use when you use comprehensive instructions that might conflict with the native system prompts.

   3. Normal - default of the system

2. Temperature-
   The temperature value is defined as the level of creativity.
   A low temperature (near 0) makes the AI stick to the most likely answers, so its replies are very predictable and “safe.”

   A high temperature (up to 2) lets the AI take more risks, using less common words or ideas, so its replies can be more surprising and creative.
   

3. Number of chunks in search results-

   When the scope of content exceeds the model's context limit (e.g., querying all company data), the system performs a search and retrieves the most relevant chunks of content for the AI model.
   
   Returning more chunks provides a broader context but may overwhelm the model, potentially causing confusion. Conversely, returning fewer chunks yields more precise answers but may omit important context. If, for example, you are looking to summarize better, to have a bigger number of chunks, and if you have a specific question, it is better to have a smaller number of chunks.

4. Question Source Verification
   Sufficient keywords- in the case that it is not a summary question, the system will verify that all main keywords exist in the document, or else it will return that the answer is not based on the document. This is done only if the user has enabled “Question Source Verification”. The default is defined in the user settings and can be changed in the Advanced Options of the General Chat.
   This verification is done before sending content to the AI

5. Answer Source Verification

   Checks answers with company data and rewrites answers if necessary.
   ​https://agatsoftware.atlassian.net/wiki/spaces/VA/pages/4534468610/Answer+Source+Verification#Modes

6. Maximum Effective Context Tokens
   The maximum number of tokens sent to the model to generate an optimal response.

   • If the "collection" option is selected and the entire collection fits within the Maximum Effective Context Tokens limit, the system sends the complete collection.

   • Otherwise, the system performs a search of chunks to retrieve a specified number of chunks (default: 30) from the search results. In this case you should see in the auditing “User similarity search=True”

   Selected Documents from search
   Here’s a rewritten version that matches the select_partial_documents logic more accurately:

Search Processing Logic

The system first analyzes the returned chunks to identify their source documents and groups them by document ID. It then evaluates whether any documents can be promoted from chunk-level results to full-document results without exceeding the available token budget.

1. Selected Documents from search
   If the set of unique source documents can fit within the Maximum Effective Context Tokens limit, the system sends all of those documents as full documents.
   This is classified as select_documents.

2. Dominant Document from search
   If the full set of documents does not fit, the system checks whether one document dominates the search results.
   If at least 60% of the relevant top chunks belong to a single document, the system sends that document only and uses the retrieval/refine path.
   This is classified as dominant_document.

3. Select Partial Documents
   If neither of the above applies, the system selects the most relevant documents that can fit within the token budget, promotes some of them to full documents when possible, and keeps the remaining results as chunks. used PrecentChunksForSelectPartialDocuments config (15% by default)
   This is classified as select_partial_documents.

4. Stuff
   If no document promotion is possible, the system falls back to sending the retrieved chunks directly, limited by the maximum chunk budget.
   This is classified as stuff.

When a single document is selected, the system will send the entire document if it fits in the context window

Question analysis - AI Router

The system has a component that analyzes the prompt and the content sent to the AI model, which is referred to as AI Router

The router has the following parts:

Question analysis- Checks what type of question this is in order to decide if the entire document is required and if keyword verification is needed. For example, it will check if the prompt is a summary question. If it is a summary, it sends all the documents and does not check for Sufficient keywords.

Using Tools

You can use tools defined by the admin by chosing them from the list.

You can also chose a general AI Agent chat and simply write a promt using the relevant tools for example “Createa an image of a dog and send it to me by email”