ExoMind - Exo Insights

14 min

Feature Overview

Exo-Insights, also referred to as Exomind, is a generative AI-powered feature that transforms your chatbot from a flow-based static system into a dynamic, knowledge-based assistant.

Instead of manually building a conversational flow for every possible question, you can create a "knowledge base" by uploading documents, websites, and other resources. The bot then uses this knowledge base to understand and answer a wide range of free-form user queries in a natural, human-like manner.

This feature is designed to handle the "long tail" of user questions that are impractical to cover with individual, pre-defined flows. By training Exomind on your company's support pages, product manuals, policy documents, or FAQs, the bot can intelligently search, synthesize, and present relevant information, complete with source links for verification.

Key capabilities of the Exo-Insights module include:

Knowledge Base Creation: Train the bot by uploading various document types (PDF, DOCX, XLSX, TXT, etc.) or by providing website URLs.

Natural Language Understanding: Allows users to ask questions in a conversational, free-flow manner without needing to use specific keywords.

Source Referencing: The bot can provide a direct link to the document or webpage from which it sourced the answer, enhancing user trust.

Priority Management: Assign priorities to different documents to control which source the bot prefers when information overlaps.

Customizable Responses: Adjust settings like "Temperature" to control the creativity of the bot's answers and "Strictly Answer From Knowledge Sources" to prevent it from responding outside its training data.

How to Use the Feature

This guide explains how to enable Exomind, build its knowledge base, and manage the content it uses to answer user queries.

Enabling and Configuring Exomind

From the main console menu, navigate to Settings -> Solution Settings.

Enable the Enable PDF Searcher toggle.

From the Choose PDF Searcher Model dropdown, select ExoMind.

Configure the additional settings:

Show Source Link: Check this box to have the bot include a link to the source document with its answers.

Temperature: Adjust this slider to control the randomness of responses. 0 is very factual and deterministic, while 1 is highly creative. A setting around 0.7 is common.

Strictly Answer From Knowledge Sources: Enable this to prevent the bot from answering questions if it cannot find relevant information in its trained knowledge base.

Click Save.

Note: If the feature might not be available on your bot, please reach out to your respective Account Manager to get it enabled.

Best Practices: Using Exomind with AI Agent (Recommended)

While Exomind can be used independently, we strongly recommend using Exomind together with the AI Agent (Bot Mandate) for most use cases.

Why Use Exomind with AI Agent

Using Exomind inside an AI Agent node allows you to:

Define a custom prompt to control how the bot answers.

Shape tone, structure, and verbosity of responses.

Combine retrieved knowledge with the conversation context.

Build more flexible and controllable GenAI experiences.

Using Exomind without AI Agent relies on default behavior and does not allow custom response prompting.

Recommended Setup Flow

For most use cases, follow this order attached a GIF below:

Enable Exomind

Go to Enable PDF Searcher → Exomind

Upload and train your documents or FAQs

Enable AI Agent (Bot Mandate)

Go to Settings → Solution Settings

Enable AI Agent

You will see 3 AI Agents created by default.

Connect Exomind Inside AI Agent

Navigate the AI Agent studio from the LHS navigation bar bu clicking Agents.

Open a sub-goal node called "FAQ Node" in AI Agent Studio, while you can Delete the other AI Agent node if not required.

Inside the FAQ node, check the default prompt and update it as per your needs.

Click on the Home icon in the Prompt editor, enable the AI Agent toggle.

Go to Action and select to update the Action name.

Select the Action Type = Exomind, a default Exomind of the same bot will automatically connect to the AI Agent node.

This setup allows you to build flexible, prompt-controlled, GenAI-powered knowledge bots.

When to Use Exomind Without an AI Agent

You may continue to use Exomind without the AI Agent if:

You prefer a simple, out-of-the-box knowledge base experience

You do not need custom prompting or advanced conversational control

However, for most customer-facing and demo use cases, Exomind + AI Agent is the recommended approach.

Building the Knowledge Base

Navigate to Workflow -> Exomind. This will take you to the knowledge base management page.

Click on Add New Source.

Choose the type of source you want to add:

For Documents (recommended for best accuracy): Select "Upload Documents" and drag-and-drop your files (e.g., PDF, DOCX, TXT).

For Websites: Select "Add Web Page" and enter the URL. Refer to the best practices provided below.

After adding the sources, they will appear in the table with a status of "Not trained." You must select them using the checkbox and click Start Training.

The Status column will update to "Training" and finally "Trained" once the process is complete. If it shows "Training Failed," you can hover over the status to see the reason.

Managing Sources

Assign Priority: You can set a numerical priority for each document. When information conflicts, the bot will prefer the source with the lower number (1 being the highest priority).

Delete: Select one or more sources and click the Delete button to remove them from the knowledge base.

Download/Copy: For documents, an icon to Download the original file is available. For web pages, there is an icon to Copy link.

Common Examples and Use Cases

HR Policy Bot: Upload the company's HR handbook, leave policy documents, and benefits guides. Employees can then ask natural language questions like, "How many days of sick leave do I get per year?" or "What is the company's policy on remote work?"

Product Support Assistant: Train Exomind on technical manuals, user guides, and troubleshooting articles. Customers can then ask questions like, "How do I reset my device to factory settings?" or "My screen is flickering, what should I do?"

Website Content Search: Provide the URL of your main website or support section. The bot will crawl the content, allowing users to ask questions about anything on the site directly within the chat window.

We have a sample PDF file with Q&A, and have created a Credit Card Related FAQ topic for reference:

Credit Card FAQ

Best Practices for Creating Documents for ExoMind

This guide outlines how to create, structure, and maintain documents for optimal performance of Exomind, our RAG-based Knowledge AI system for Chatbots. Following these standards ensures higher ExoMind / RAG accuracy.

1. Document Structure

One Document: Instead of multiple documents, have a single collated document for one ExoMind (in PDF format)

Main Heading: Should clearly reflect the product or topic (e.g., Loan, Savings Account, Credit Card). Each Heading should start from a different page.

Sub Heading (Optional): Divide content into logical sections for easier navigation. Avoid very deep hierarchy. Keep it two levels only (Heading and Sub-Heading). Each Sub-Heading should also start from a different page.

Separate Sub-Heading for Dynamic Content: Create a separate sub-heading for content which is Dynamic (Changes frequently)

Sub Heading Type	Description	Update Frequency
Static Information	Core product details, eligibility, features, general FAQs	Rarely (More than 2 months)
Dynamic Information	Offers, rates, operational updates, announcements	Frequently (Less than 2 months)

Question / Answer Format (FAQ Format): Inside each Heading/Sub-Heading, format the content into Question/Answer format only

Questions can be one or more with a common answer.

There will be unique “Question(s)/Answer” pairs in any Heading / Sub-Heading

Inside each sub-heading, the Questions must be Mutually Exclusive and Collectively Exhaustible.

Single-Page FAQ: Keep each FAQ on a single page — the answer must not spill over to the next page (Half answer on one page and the remaining answer on the next page)

Avoid Multi-Topic Pages: Each page should cover a single, well-defined topic. Only very similar questions should be on the same page. Whenever there is confusion, it is better to move to another page. Each page of the document is one RAG chunk

Consistent Formatting: Maintain the same font, header hierarchy, and spacing across all documents.

2. Content Writing Guidelines

Do not have the same question twice with different answers in the document. This is most important to ensure ExoMind accuracy.

The answer must be comprehensive. It should cover everything, including any potential follow-up questions. Two different questions can have some sections in the answer which are common. No answer should be more than 3 paragraphs.

Each Question must be complete.

It should make sense without knowing the headers, sub-headers, etc. Frame questions as users will ask them. Focus on keywords.

There should be an initial question which helps explain the product in simple language.

Ensure every “Question(s)/Answer” pair is self-contained and does not rely on another pair

The document should be written in the tone which you want the BOT to speak. We suggest keeping the tone friendly and helpful.

Avoid jargon as much as possible. If you need to use any jargon, define the jargon within brackets.

If you are using any Acronym, add the full form in brackets each time you use the Acronym (not just at the beginning)

Remove Product-Specific Tags: Exclude internal terms like “JIFFYearn” or “#JIFFYearnCash” unless necessary for comprehension. Example: ❌ “How to open a JIFFYearn or #JIFFYearnCash time deposit?” ✅ “How to open a time deposit in JIFFY Bank?”

Avoid cross-references like “see above” or “refer below” or “same as above”

Do not use different terms for the same thing. Avoid Inconsistent Terminology (“Fixed Deposit” / “Time Deposit”).

Either use both words in all places, or use one single word at all places and one FAQ where it is explained that both these words mean the same thing.

Do not add a screenshot of Text

Do not add any image (any non-searchable text) or scanned PDFs

3. Document Naming, Priority and Sanity Testing

The document/PDF should be named with a version or date.

Use consistent naming if multiple documents have to be added in an ExoMind

Examples: Loan_FAQ_25_Nov_2025.pdf, CreditCard_20_Nov_2025.pdf

Use “Priority Setting”, if there are multiple documents with different confidence levels

Note: We recommend using a single PDF document only for ExoMind.

Internally maintain version history when updating FAQs to ensure traceability.

Test for Accuracy: After each upload, test key FAQs and confirm the responses get pulled from the correct section.

4. Continuous Improvement Cycle

Regularly review and refine documents using analytics and failed queries.

Frequency	Activity
Monthly	Review top failed queries and restructure content if retrieval gaps persist.
Quarterly	Refresh dynamic data (rates, offers, contact details).
Ongoing	Maintain naming consistency and terminology alignment across all sources.

5. Testing and Debugging Protocol (Post Go Live)

When sharing or analyzing test results, always include:

Timestamp and Date of the test

Full conversation screenshot (for debugging context)

If an Issue is Found:

Go to Home in the test interface.

Re-ask the same query.

Confirm if the issue still persists.

If yes, capture full logs and share with the Exotel Support team.

6. Common Mistakes to Avoid

Mistake	What Goes Wrong	Impact on Bot Accuracy	Recommended Best Practice
Uploading scanned or image-based PDFs	Text cannot be extracted or is partially unreadable	Poor answers, missing responses, higher hallucination	Upload only text-based PDFs or run OCR and validate text before training
Training directly on poorly structured websites	Navigation menus, banners, footers, and mixed content get indexed	Low relevance, incorrect section matching	Scrape the website and convert content into clean, structured Q&A documents
Forgetting to click Train Sources	Uploaded data is never ingested by the model	Bot ignores new or updated content	Always trigger Train Sources after uploads or updates
Uploading long, multi-topic pages	Multiple topics exist in the same chunk	Reduced retrieval precision, wrong answers	Split content into smaller, topic-focused documents or Q&A pairs
Overlapping or duplicate content across sources	Same answers appear in multiple documents	Ranking confusion and inconsistent responses	Deduplicate content and keep a single authoritative source
Missing content priority or update strategy	Older and newer versions coexist	Outdated or conflicting answers	Remove deprecated documents and retrain after updates
Expecting transactional or workflow behavior	Model is asked to perform actions or logic	Incorrect expectations and failed use cases	Use Exomind only for informational Q&A; use APIs or AI Agents for workflows
Assuming RAG will fix poor-quality data	Ambiguous or poorly written content is ingested	Hallucinations and low-confidence answers	Clean, normalize, and validate content before training