Here’s how it works:

1. LLM Preprocessing (Outside the Einstein Search Quota)
   Before any searches, an LLM:
       • Summarizes the case
       • Extracts key metadata like:
             – Products/features
             – Region or account
             – Error types and impacts
             – Timelines and dependencies
             – References to past tickets or changes
2. SOQL Filtering (Free or Token-Light)
   Using this structured metadata, we run targeted SOQL queries to narrow down candidates:
        • Only look at cases with the same product, error category, or client type
        • Optionally, generate SOQL dynamically using an LLM if sequence logic is involved
3. Einstein Search—Only Where It Counts
   Now that we’ve narrowed the pool from 100,000+ to 50–100 cases, Einstein Search is applied for deep semantic matching—but only within that refined set.
4. Optional LLM Final Ranking
   Lastly, a lightweight LLM pass ranks the top 3 results based on contextual relevance.