Building trust in AI: Implementing AI-powered semantic search for our Customer 360 data
About the Session
Enterprise executives need instant answers about customers, but critical data remains fragmented across multiple systems—Salesforce, Gainsight, Zendesk, and more. Manual data aggregation consumes hours of valuable time and creates costly delays in strategic decision-making.
We built AccountIQ to eliminate manual aggregation through natural language queries. Ask "Show at-risk EMEA accounts with high ARR—what patterns emerge?" and receive accurate, verifiable insights in under 300ms.
The system automatically consolidates data from multiple sources and understands semantic intent - recognising "at-risk" concepts beyond exact keyword matching. Unlike traditional SQL, which requires intimate knowledge of schemas and relationships, our hybrid search architecture understands business language naturally.
Most critically, we solved the AI hallucination challenge: our LLMs analyse only facts retrieved through vector search, never inventing information that executives cannot verify. Trust and accuracy are paramount in enterprise decision-making.
Key Takeaways for Implementation
- Hybrid search architecture (50/50 vector-keyword) consistently outperforms pure semantic or keyword approaches
- Search-first patterns prevent AI hallucinations by grounding LLM responses exclusively in retrieved facts
- Configuration-driven YAML schemas enable rapid iteration and scaling without code refactoring
- Modular end-to-end architecture ensures scalability and robustness for enterprise deployments
This is a story for data engineers and platform architects evaluating RAG implementations, AI/ML practitioners building enterprise solutions, and forward-thinking teams ready to embrace the future of data engineering and AI-powered analytics.
Radovan Bacovic
Staff Data Engineer, GitLab
About the Speaker
Radovan Bacovic is a Staff Data Engineer at GitLab, living, enjoying and coding in Novi Sad, Serbia.
Part of the Data gigants ambassador program: Snowflake Squad and dbt spotlight members.
An experienced data engineer and “wanna-be” is the best bad conference speaker. Forever eager to discover new data technologies in the fast-changing environment. He armoured himself with a profound application development background in large international companies around the globe. Strongly advocated for the open-source community and open-core approach.
Without any doubt - a fervent data geek delighted to share his long mileage and experience with a broader audience.