Standard Retrieval-Augmented Generation (RAG) pipelines encounter fundamental performance ceilings—termed Inference Scale Failure—when applied to large-scale unstructured corpora comprising millions of pages. These archives, typified by historically significant, high-volume collections where traditional keyword-based search has long proven inadequate, introduce compounding challenges: OCR noise, inconsistent metadata schemas, broken cross-page formatting, and redaction artifacts that degrade both ingestion fidelity and retrieval precision. This paper proposes a distributed, parallelized ingestion engine that overcomes these barriers through three principal innovations: (1) asynchronous, multi-worker extraction with noise normalization, (2) Semantic Boundary Detection for context-preserving chunking in place of fixed-size splitting, and (3) hierarchical sharded vector storage enabling sub-second retrieval across indexes exceeding three million vectors. Complementing the core architecture, the system incorporates Hybrid Retrieval—combining dense vector search with BM25 sparse retrieval—and Automated Entity Extraction at indexing time, enabling compound queries that combine named entities (persons, organizations, locations, dates) with conceptual semantic similarity. Empirical evaluation on the Enron Email Corpus (~517,000 documents), the Panama Papers dataset (~11.5 million documents), and a long-form narrative test set derived from Project Gutenberg (~50,000 volumes) demonstrates that the proposed pipeline delivers a 54× throughput improvement over sequential baselines, P95 query latency below 220 ms on a 3-million-vector sharded index, and semantic recall improvements of 23 percentage points at k=10. Exact-match entity recall improves by 57 percentage points over vector-only retrieval when hybrid BM25+vector scoring is applied. The architecture is designed to be horizontally scalable, cloud-agnostic, and fully reproducible using open-source components.
Brent Allen Jensen (Fri,) studied this question.