How Lucy Built a Source-of-Truth Retrieval Layer for 2.8M+ SEC and DART Filings with Qdrant - Qdrant

This post details Lucy's implementation of a source-of-truth retrieval layer for 2.8M+ SEC and DART filings using Qdrant. It highlights the use of Qdrant's payload filtering for precise search constrained by metadata (company, form type, date range, document ID), hybrid search (dense + BM25 with RRF) for combining semantic and lexical matching, and TurboQuant (4-bit for SEC, 2-bit for DART) for significant memory footprint reduction with minimal impact on retrieval quality. The architecture involves a router to identify the target filing, followed by parallel retrieval lanes (narrow, broad, XBRL, companion) that execute hybrid queries against Qdrant. The system achieved 94.7% Recall@10 on FinanceBench and a 0.160 gain in RAGAS Answer Accuracy compared to general web search context. Qdrant averaged 32ms per search with p90 of 82ms.