How Chat Answers are generated

Retrieval-Augmented Generation

The YUKKA News Assistant Chat is using Agentic Retrieval-Augmented Generation (RAG) setup, that enhances an underlying Large Language Model by combining its language abilities with real-time access to dedicated data sources via YUKKA platform APIs.

An agent helps the News Assistant decide whether to answer from the model or fetch specific information from various tools that leverage proprietary signals, ensuring the most relevant source is used. This setup reduces hallucinations by grounding responses in accurate, up-to-date data rather than relying solely on the model’s internal knowledge.

Tool Selection

Based on the input prompt, the agent analyzes its content and automatically selects the most appropriate tool to generate the most relevant and accurate response. This ensures that each query is handled using the capabilities best suited to its context and intent:

• Keyword Search: This tool is a news search engine that delivers the latest updates on any topic by using keyword or similarity-based searches across real-time and 12-month archives.
• Scores: This tool resolves user questions about scores by identifying the respective YUKKA score (More about YUKKA Scores)
• News Stories: An entity-based search engine that provides the latest news stories about some entity such as company, organization, person, product, city, country or currency.
• Events Search: This tool is a news search engine that provides news about events or topics relevant in the financial and ESG domains. (More about YUKKA Events)
• News Sentiments: A search engine that provides the latest sentiment information as per referenced entity Type.
• Knowledge Graph: This tool answers questions about entities by pulling together reliable data on entities like company names, financial identifiers, ownership structure, locations, industries, stock listings, and employee numbers. 

Large Language Model Foundation

When you ask a Question to the YUKKA News Assistant, here's what happens in the background to understand and respond in the best possible way.

1. Understand the Question
   The agent first figures out what the user is asking and what the question is about (e.g., a product, person, or topic).

2. Match to the Right Information
   It looks for matching data or content—this might include news, documents, events, or other stored information—and identifies the right entity (like the right product or person) based on the question.

3. Pick the Right Tool
   Depending on the type of question, the assistant chooses the best tool to answer it—this could be a search to internal APIs, a database lookup, or a graph query (like Neo4j).

4. Generate the Answer
   The selected tool retrieves relevant news or data, answers user question based on them with clear, human-friendly response and return to the user through the agent.

5. Suggest Next Steps

   After answering, it may offer follow-up questions or suggestions to keep the conversation going or help the user explore more.

The YUKKA News Assistant uses OpenAI's GPT-4o, accessed via API through OpenAI’s U.S.-based data centers. The model intelligently selects tools and generates responses based on the user's input.

This service is currently in beta. To protect your privacy, please avoid sharing confidential information.