Oracle Bets on Semantic Search Without Generative AI

In a move that challenges the current trend toward large language models (LLMs) in enterprise search, Oracle is offering a new system called Trusted Answer Search. This technology delivers reliable results by querying a curated set of approved documents using vector-based similarity matching rather than generative AI.

The system works by allowing enterprises to define a “search space” consisting of reports, documents, or application endpoints paired with metadata. When a user submits a natural language query, the system deterministically maps it to a specific match document and returns a structured outcome—like a report, URL, or action—rather than generating free-form text.

Addressing Enterprise Concerns

According to Tirthankar Lahiri, SVP of mission-critical data and AI engines at Oracle, this approach addresses growing enterprise needs for more predictable search results that provide auditability for compliance purposes. Independent consultant David Linthicum echoed this sentiment, noting that the technology would appeal to organizations in regulated industries like finance and healthcare where consistency is paramount.

Trade-offs and Challenges

While Trusted Answer Search avoids LLM inference costs by shifting spending toward data curation and governance, experts caution about potential trade-offs. Robert Kramer, managing partner at KramerERP, points out that CIOs will need to budget for ongoing maintenance of the curated content.

As the volume of trusted documents grows—potentially including regulatory updates or market reports updated frequently—the risk of serving inaccurate information increases, warned Scott Bickley, advisory fellow at Info-Tech Research Group. He noted that even with curation, similar language appearing in different contexts can lead to plausible but incorrect answers.

Dynamic Content Approach

To address these concerns, Oracle’s system treats trusted documents as parameterized URLs that pull content dynamically from live data sources rather than relying solely on static repositories. This allows it to generate responses from enterprise applications, APIs, or regularly updated web endpoints.

While this approach helps reduce content churn, Linthicum emphasizes that maintaining accuracy still requires disciplined governance and feedback mechanisms, especially with thousands of targets where semantic overlap can occur.