Get a taste here of some of the learnings, insights, and key takeaways we discussed with founders & top management on the August 30th roundtable.

Want to engage in “real-time discussions on GenAI”, and hear more from your peers? – Join our WhatsApp GenAI community by subscribing here.

1/ Things that are being developed internally by Capria:

• Internal QA Bot: This bot queries around two thousand documents in Capria’s internal repository, providing answers and insights.
• IC Memo and Alert Tool: This was created to streamline the generation of Investment Committee (IC) memos and alerts. It offers suggestions for improvement, aiding in the creation of effective and informed investment documents.
• Transcription and Analysis Tool: Transcribes conversations from one language to another and identifies the five most relevant topics discussed to help address needs more effectively.
• Natural Language to JSON Conversion: This project involves converting natural language requests into JSON or other formats, enabling interaction with APIs. This enables users to communicate their needs in a conversational manner.
• Interacting with Databases: This aims to allow natural language interaction with databases, such as querying data without writing complex SQL commands. Users can ask questions in plain language, and the tool retrieves relevant data from the database.

2/ Some taste of our own learnings along the way:

• Language Models for Semantic Information Retrieval: Modern language models are highly effective in semantic information retrieval. Comparing them to human capabilities, these models perform well in answering questions and providing relevant information. These models excel when used in Q&A tools, returning accurate and often correct responses.
• User Feedback as Improvement Tool: It provides key user feedback in enhancing the system’s performance. By allowing users to offer feedback through options like thumbs-down or detailed comments, the system can be refined and corrected over time (iterative process for continuous improvement).
• Deploying Language Models in Production: While comprehensive testing is valuable, it might not be feasible to allocate exclusive time for it. Deploying the system and collecting user feedback in a real-world setting helps refine the system and reduces the gap between development and deployment.
• Choosing Between Self-Deployed and API-based Models: We opted for the API service due to its effectiveness and cost-efficiency. We recommend starting with an API, especially for small to medium-scale organizations, and considering self-deployment later based on evolving needs.

3/ Some taste of the challenges we faced:

• Conflicting Information and Prioritization: With a growing number of documents, the system encounters conflicting information on certain queries. For instance, asking for an office address could yield multiple documents with different addresses. The challenge is that the system can’t inherently prioritize one source over another. This can lead to incorrect answers being provided without clear indications of their inaccuracy.
  • Solution # 1 – Manual Verification and Date of Modification: One approach to solve these issues is manual verification, wherein conflicting information is identified and corrected manually. However, this becomes impractical when dealing with a large volume of data. We propose a more automated solution: associating a date of modification with the information. Bypassing the date of modification along with the data, the system can give priority to the most recent and updated information, logically assuming it to be correct.
  • Solution # 2 – Relies on the following principle: “Every document can be queried in a finite number of ways by an information source”. In this method, we pass an information source through a Language Model and generate a comprehensive list of potential questions. Subsequently, a purely mathematical function is employed to assess the similarity between the query and the automatically generated questions in an embedding space, determining their relevance. To further enhance accuracy, we can supply both the generated questions and the query to the Language Model for evaluating relevance. By doing so, we can be confident that out-of-context hallucinations are reduced to 0, as the questions and their respective answers are pre-determined.

4/ Some taste of what other founders & top management shared:

• Importance of Data and Chunking in AI Models: Standard models and chunking methods may not always yield satisfactory results with vector embeddings. The accuracy and results of AI models can significantly improve when the data set is changed and chunking is optimized. Properly chunking data and ensuring relevant context can greatly impact the accuracy of vector embeddings.
• Tailoring AI Solutions for Specific Needs: Building AI solutions tailored to specific needs can yield unexpected benefits. Developing tools like a resume builder with unique features such as Master JD (Job Description) and Master Resume can greatly improve accuracy and reduce manual effort.

Other topics discussed in the live session:

• How to come up with hyper-personalization
• Scale and challenges in vectorization
• Balancing affordability and accuracy
• Iterative process and AI collaboration
• Contextual chunking for improved accuracy
• Tailoring AI solutions
• Language and grammar challenges
• Leveraging AI models for cost savings

Don’t miss the opportunity to hear your peers sharing insights on these topics and more. You’ll have the opportunity to engage in techy GenAI discussions and also share your learnings, challenges, and any doubts you may have. Join our WhatsApp community by subscribing here.