In today’s technology-driven world, we are surrounded by an overwhelming flow of information. But within this vast stream, is it really easy to tell what is true and what is false? From Facebook news feeds and WhatsApp and Telegram group chats to news apps, false news, rumors, and half-truths are everywhere. For the millions of Bengali-speaking people, this is a major challenge. Without a reliable system for verifying information, the consequences can be devastating.
It is in this era of overwhelming misinformation that “Khoj” emerged, the first and most comprehensive AI-powered fact-checking platform in the Bengali language [1], which began its journey on September 1, 2024.
Whether it is text, images, historical events, or rumors spreading virally online, Khoj can become your reliable companion for verifying information, right in your native language. It was created with the goal of restoring truth in the digital age, making it a potentially essential tool for improving digital literacy.
Imagine coming across a piece of news online one day that is not reported by any major media outlet, yet the number of shares and comments on Facebook and WhatsApp makes it seem genuine:
“Billions of taka have been looted from a particular bank, leaving thousands of people financially devastated.”
In such a moment of uncertainty, Khoj can be your trusted source of verification. Within seconds, its AI will search for the original sources, compare the reliability of the available information, and present you with a comprehensive fact-checking report. It will tell you what makes the claim credible, where the information came from, and where there may be reasons for doubt. This fact-checking does not simply appear out of nowhere like magic. Nor is it the case that AI is all-knowing and can therefore fact-check everything for you. There is sophisticated technology working behind the scenes.
If you want the process in a nutshell, here it is. First, Khoj searches the web, social media posts, news archives, and multimedia related to the claim. It then evaluates the reliability of the sources, analyzes text and images while considering their context, and finally provides a clear verdict with the supporting sources, whether the claim is true, misleading, or false. Now, let us go a little deeper and explore how Khoj actually works.
How Khoj Works
The core strength behind Khoj is a smart architecture developed by the Khoj team. It works together with AI to retrieve information from reliable sources and present it in a way that is easy for anyone to understand. Calling it merely a search engine would be an oversimplification. It is an intelligent process designed to make every effort at every stage to produce the most accurate results possible. Let us take a detailed look at how it works.
(1) Curated List of Trusted Sources
We carefully researched and compiled a list of more than 316 Bengali and English websites. These are sources where reliable information on current affairs, politics, economics, celebrities, science, religion, health, and other topics can be found. The list includes major news portals, fact-checking websites, scientific journals, and health-related platforms. Why was this necessary? Because if the source of information is unreliable, the entire process becomes ineffective. This carefully curated list significantly reduces the risk of “garbage in, garbage out.” As a result, our system attempts to ensure that every result comes from recognized and verified sources, strengthening the reliability of the information presented to users. The goal was to build a secure library where only authentic and validated sources are retained.
(2) Query Optimization: The First Step of AI
When you ask a question, such as “Was there vote rigging in the DUCSU election?”, that is when our AI begins its work. It transforms your simple question into precise and effective search keywords. This process is known as query parsing.
For example, your question could be transformed into:
- “DUCSU election 2025 vote rigging”
- “DUCSU election vote fraud”
- “allegations of irregularities in DUCSU election”
- “DUCSU election result controversy”
The AI then adds the names of relevant and reliable news organizations and fact-checking websites to these keywords, making the search more targeted and precise.
Here, the AI is specifically instructed through a system prompt to identify the most effective search keywords. This significantly reduces irrelevant results and increases the likelihood of finding accurate information. This step is one of the key aspects that distinguishes Khoj from conventional search tools. In simple terms, it directs your question toward the most relevant information with precision, saving both time and effort.
(3) Web Crawling and Filtering
As discussed above, in the first stage, Khoj processes the user’s question through query parsing and transforms it into smaller, more effective keywords. In the next stage, the search begins using the curated source list that we established at the beginning.
For this purpose, Khoj uses a Web Crawling Pipeline that scans the HTML DOM Structure of each website and collects the necessary content, including the title, meta tags, and main body text. For those unfamiliar with web crawling, it simply means navigating websites and retrieving data from them. The DOM structure refers to the underlying structure of text and tags within a web page.
The documents collected through crawling are then evaluated using a Weighted Scoring Algorithm. In this algorithm, each source is evaluated using different weights. In simple terms, these weights indicate how important one factor is compared to another.
Three primary metrics are used to calculate the weight assigned to each source:
a) Recency Score: This score indicates how recent a piece of content is. The system distinguishes between newer and older information based on the content’s timestamp.
b) Authority Score: This score is used to measure how trustworthy a website is. A score is determined by analyzing factors such as the website’s Domain Authority (DA), PageRank, and trust signals.
c) Keyword Match Score: This metric measures where the keywords appear and how relevant those occurrences are to the actual query. It uses TF-IDF, or Term Frequency-Inverse Document Frequency, to determine how closely the keywords match the title and body text.
Weighting Rule: A keyword match in the title is given greater weight, while a match in the body text receives comparatively less weight. By combining all the metrics, a Composite Score ranging from 0 to 1 is calculated for each source. In other words, this score represents the final value obtained by combining all the relevant factors.
Finally, the 8 to 10 sources with the highest scores are stored in JSON format. For those unfamiliar with JSON, it is a data format that can be easily read and processed by computers. The filtering process is designed to be precise enough to resemble finding a needle in a haystack. The result is that a large amount of irrelevant information is filtered out, leaving only the most reliable, recent, and relevant data.
(4) Fallback Search: Leaving No Gaps
As we mentioned earlier, Khoj retrieves information for users from its curated sources. But what happens when a claim is extremely new or unusual, and no relevant information can be found among the sources listed in Khoj? In such cases, Khoj does not simply stop. Instead, it immediately falls back to a specialized tool, the Tavily API. It essentially searches across the web while retrieving information only from reliable and authentic websites.
This process also includes a smart authority scoring system that helps ensure that fake or irrelevant links do not make their way into the results. On top of that, Khoj has 16 separate API keys built into the system, which serve as backups for one another. If one stops working, another takes over. As a result, Khoj can continue operating even when it encounters rate limits or other usage-related issues.
In short, this acts as Khoj’s robust safety net, ensuring that there is always a way to investigate the truth, no matter how difficult or complex the claim may be. According to the system’s reported performance, this fallback mechanism contributes to an accuracy rate of approximately 95%, which is claimed to be higher than that of many international platforms.
(5) Report Generation: AI Creativity and Accuracy
Once all the information has been collected, Khoj’s AI does not simply stop there. It organizes the findings into a clear, easy-to-understand, and reader-friendly report. For this, a three-level AI fallback system is used, ranging from the primary level to the tertiary level, so that the possibility of errors is reduced as much as possible.
The report includes:
- What different sources are saying
- Where disagreements exist
- Whether the claim is ultimately true, false, or disputed
This is where the real strength of the AI becomes visible. First, it optimizes your question in a way that helps retrieve the relevant information. It then organizes the results into an easy-to-follow narrative while taking the cultural background into consideration.
Each report includes source links, allowing you to verify the information yourself if you wish. As a result, the experience feels almost like having a conversation with an expert. It is simple, reliable, and highly convincing.
Specialized Features: What Makes Khoj Stand Out
To make Khoj a complete platform for seeking the truth in the Bengali language, several additional useful features have been incorporated:
Mythbusting: In the digital age, false news and rumors have become part of our everyday reality. Most of the time, people stop after simply finding out whether a particular claim is true or false. Khoj, however, takes a somewhat different approach.
One of Khoj’s most distinctive features is that it does not simply provide information. Instead, it explains the subject in a conversational manner. Suppose you come across a superstition or a scientific claim. The AI will explain it to you almost like a friend. It will first provide the context, then present scientific evidence, and finally explain the misconception in a narrative and easy-to-understand way.
The greatest strength of this feature is that it does not merely protect people from rumors. It also helps develop critical thinking skills in the long term. That ability creates a form of digital protection that can be highly valuable in the information conflicts of the future.
Liberation War Corner: Even today, various misconceptions, distorted versions of history, and rumors surrounding 1971 continue to circulate, potentially misleading the younger generation. In this context, Khoj has introduced a unique feature called the Liberation War Corner.
Here, the AI works like an experienced historian. It does not merely present information. Instead, it explains everything from historical context, timelines, genocide statistics, war crimes, and important personalities to geographical details and cultural impacts in a simple and engaging narrative style. What makes this feature distinctive is that it has been developed using the Retrieval-Augmented Generation (RAG) approach. It draws upon numerous historical books, government documents, diaries, and authoritative archives related to the Liberation War. As a result, the information is not merely an AI-generated assumption. It provides evidence-based explanations retrieved from verified sources.
This makes it easier to identify the various false narratives, historical misconceptions, and rumors surrounding the Liberation War. The aim is to ensure that the true story of our struggle, sacrifice, and victory is not obscured by distortion.
- Multimedia Verification: Khoj can verify not only written information but also the authenticity of images. Its capabilities range from detecting AI-generated content to finding the original source of an image through reverse image search. For text, Khoj uses AI detection and plagiarism checks to help identify deepfakes or manipulated content. With advanced API integrations, Khoj is now equipped for the multimedia-driven information environment of the digital age.
Manual Fact-Checking: A Strong Human-AI Collaboration
Alongside its AI-based features, the Khoj team regularly performs manual fact-checks on various rumors circulating online. The objective is simple: to prevent misleading or false information from spreading easily. In this way, the speed of artificial intelligence and the analytical capabilities of humans work together, turning Khoj into a reliable standard for information verification.
Limitations
No system is completely perfect. Regardless of how extensive the source list may be, information from sources outside that list may be missed. This creates the challenge of source bias. There may also be slight delays when real-time updates are unavailable. AI-generated summaries can occasionally contain subtle errors, which is why it is always important to verify the original source links.
Conclusion
Khoj is more than just a platform. It is a new initiative aimed at improving digital literacy in the Bengali language and helping people identify misinformation. Its hybrid pipeline, combining AI-powered search, source selection, and summarization, demonstrates how information can be made easier to verify. We hope Khoj will continue to assist users and gradually create a greater impact. Whenever you are searching for the truth, Khoj will be there beside you.
You can join Khoj on this journey in the pursuit of truth: khoj-ai.bd
Reference
[1] Fuad, Fardin-E-Hasan. “ফ্যাক্টচেক করবে এআই, ৬ তরুণের উদ্যোগ.” Prothom Alo, 30 Nov. 2025, www.prothomalo.com/lifestyle/sb6clatm04.