Two-Pass LLM Processing: When Single-Pass Classification Isn't Enough

This article discusses a pattern where a single-pass LLM classification approach fails to capture the relationships and context between items, leading to incorrect classifications. It proposes a two-pass architecture to address this issue.

Details

The article describes a common scenario where you need to classify a batch of items (messages, tickets, documents, etc.) using an LLM. The obvious approach is to make one LLM call per item, which works fine until it doesn't. The failure mode is that while each item may be correctly classified in isolation, the relationships between items - such as escalation patterns, contradictions, or duplicate reports - are invisible to a single-pass classifier because it never sees the full picture. The solution proposed is a two-pass architecture. In the first pass, each item is classified independently, generating per-item labels for category, urgency, summary, and suggested action. In the second pass, all items and their initial classifications are fed back into the LLM, which is then asked to find relationships, patterns, and adjust classifications based on the full context. This allows the system to identify threads of related items, raise cross-item alerts, and provide a synthesized summary.