AWS re:Invent 2025 -Fidelity Investments: Text-to-SQL for data analytics at enterprise scale-IND3323

Fidelity Investments: Text-to-SQL for Data Analytics at Enterprise Scale

Business Challenge and Approach

• Fidelity Investments observed that their users were struggling to access and derive insights from their data
• Users were relying on multiple applications, spreadsheets, and manual collaboration to piece together the information they needed
• The root cause was not the data itself, but the difficulty in accessing and querying the data
• This challenge is common across enterprises where data is scattered, complex, and requires technical SQL expertise to leverage

Fidelity's solution was to enable a "natural language to SQL" capability to democratize data access and empower business users:

• This approach brings the user's questions directly to the data, rather than forcing users to navigate complex data structures
• It reduces dependencies on technical SQL experts and enables a wider range of personas within the organization to self-serve data insights

Technical Implementation

Fidelity's text-to-SQL solution is built on three key pillars:

1. AI-Ready Data

• Fidelity followed four principles to structure their data for AI/ML consumption:
  1. Use consistent naming conventions
  2. Establish strong relationship foundations with explicit primary and foreign keys
  3. Leverage product hierarchies to enable pre-aggregated data
  4. Ensure a single source of truth

2. Semantic Layer

• Fidelity created a semantic layer to translate business language into database terminology
• This includes detailed metadata like business descriptions, example values, and relationship definitions
• An automated tool was built to generate these semantic files, with human review to ensure accuracy

3. Dynamic Workflow

• The workflow includes query rewrite capabilities to handle vague or ambiguous user questions
• Context engineering is a critical component, assembling the right information (schema, examples, entity values, etc.) to feed the language model
• Security guardrails are in place to only allow SELECT queries, not DML operations
• Iterative learning is enabled by storing validated question-SQL pairs and using them to improve the model over time

Performance and Lessons Learned

• Fidelity's current performance metrics include:
  • Execution accuracy: 93-95%
  • Result size comparison: 89-92%
  • Average response time: 28 seconds
• These metrics have improved over time through continued learning and optimization
• Key lessons learned:
  1. Use a deterministic, production-ready workflow from the start
  2. Maintain a clear, well-documented data schema to prevent interpretation errors
  3. Focus on quality over quantity of teaching examples
  4. Optimize the context fed to the language model for efficiency

Business Impact and Use Cases

• Fidelity's text-to-SQL solution has democratized data access and enabled a wider range of business users to self-serve insights
• It has reduced dependencies on technical SQL experts and allowed users to focus on their business questions rather than data access challenges
• The solution has been applied across various business units and use cases within Fidelity, empowering users to make more informed, data-driven decisions.