Market news assistant

• 

  Market News Assistant

Inspiration

The project was inspired by the possibility of combining abstract, high-level reasoning with retrieval of specific information from a large internal dataset to provide well-grounded, intelligent responses focused on a particular domain of interest (financial market news, in this case). Typically, sample implementations of agentic assistants demonstrate search capabilities using Google Search or tools like yfinance. Both approaches have their limitations. Google Search may be too broad and lack specificity, while data sources like yfinance are, by contrast, too rigid and limited in scope. The goal was to implement an approach that could overcome these limitations and provide more balanced, flexible access to information.

What it does

It is a voice-enabled agent that combines semantic database search with high-level reasoning to provide well-grounded intelligent responses about financial market news.

How we built it

They key functionality is distributed between two components. The first one is the ADK agent, which is responsible for high-level reasoning, natural language understanding and tool calling. The second one is the Elasticsearch, which is responsible for providing MCP endpoint, retrieving relevant news articles from the database and includes semantic search capabilities. The overall architecture overview is attachached as a separate image.

Tools used

• Python ADK Live Voice Agent with gemini-2.5-flash-native-audio-preview-12-2025
• Elastic Index for storing and searching news articles
• jina-embeddings-v5-text-small model for text embeddings
• ES|QL Semantic search
• Elastic Agent Builder Tools and MCP
• Google Cloud Run (deployment)

Challenges we ran into

The main challenge is to generate concise responses that are simultaneously natural-sounding and well-grounded in the specific facts.

Accomplishments that we're proud of

ADK is a very flexible framework. In this project, we demonstrated one of the dimensions of this flexibility that is not widely discussed in demo materials - the ability to integrate with non-Google tools and services widely used in the corporate world (Elasticsearch in this case). The demonstrated integration with Elasticsearch semantic search potentially allows to achieve high precision, speed, flexibility and credibility in information retrieval, which is crucial for real-time interactions with voice assistants in business context and other domains.

What we learned

The project was a great opportunity to practice end-to-end implementation of an agentic assistant involving external tools integration and semantic search in a specific dataset, which is not widely discussed in tutorials and demo materials.

What's next for Market news assistant

Generate text summaries of the news articles and charts and provide them to the user as part of the answer. Combined audio-visual presentation of the responses based on multiagent architecture.

Built With

• adk
• elasticsearch
• embeddings
• es|ql
• gemini
• mcp
• python