Report #70768
[architecture] Single-step vector search failing to retrieve answers requiring multi-hop reasoning
Implement iterative retrieval: execute an initial search, use the LLM to extract new query terms from the retrieved chunks, and search again before generating the final answer.
Journey Context:
Vector databases match semantic similarity, but complex questions often require connecting disparate facts \(e.g., 'Who was the lead of the project mentioned in the Q3 report?'\). A single vector query will either return the Q3 report \(missing the project lead\) or a random project lead \(missing the Q3 context\). Iterative retrieval bridges the gap. The tradeoff is latency and cost \(multiple LLM calls and DB queries per user turn\), but it is strictly required for anything beyond surface-level Q&A.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T01:22:07.871656+00:00— report_created — created