FinSight
Project Description
I built the backend systems for New Predictor, where I architected the web scraping pipeline, API integrations and deployment of the system. News Predictor is an AI-powered news analysis platform designed to help users quickly understand the real world-impact on markets. The system aggregates news from multiple sources, analyzes the content using LLMs, curates financial data and produces summaries together with market impact in an user friendly way
My role involved building the
- News Article curator that scraped financial news articles
- Financial data curator that pulled all related financial information about related stocks
- LLM Integration and reliability optimization that ensured all claims are verified with direct quotation
This page will go over the system I implemented to reduce hallucinations in LLM output
Chain of Thought Prompting
- How it works and Why its Used
LLMs are statistical systems that learn from vast amounts of data. Their knowledge is based on the data they are trained on which may contain errors which can lead to hallucinations- where LLMs present misinformation as fact. One of the ways to prevent hallucinations is called Chain of Thought Prompting. Chain of Thought (CoT) Prompting is a technique used to encourage LLMs to reason before arriving at a conclusion.
This is particularly useful in complex tasks like analysis. Implementing CoT prompting involves adding step by step reasoning instructions, providing examples of reasoning for previous questions. By providing concrete steps to follow, LLMs identifies the source of each claim, reduces the likelihood of generating hallucination and provides consistent feedback.
Additional Steps
Enforcing Verifiable Claims, Future Improvements
A key principle in NewsPredictor is that every claim made by the LLM must be accompanied by a direct quote from the news source. This allows users to verify information independently further bolstering the reliability of our outputs. The pipeline I build ensures LLM output link each insight to the corresponding quote and is cross referenced to confirm consistency.
Future Improvements:
- Incorporating more complex Sentiment Analysis systems would give more accurate predictions
- Adding additionally verification steps to ensure reasonings are logically sound and represent the quotes accurately.