Dev.to
6/29/2026

The original title is: "Two audits of my own knowledge graph found two unrelated silent failures"
Original: Two audits of my own knowledge graph found two unrelated silent failures
Short summary
Author discovered two silent failure modes in their Rust knowledge graph system: duplicate ingestion (52% redundancy) and silently dropped scanned PDFs (32% of corpus). Fixes involved deduplication by similarity and OCR fallback, improving evaluation scores from 58.3% to 63.9%. Practical debugging patterns for RAG and document ingestion pipeline reliability.
- •Silent ingestion duplicates inflated graph size by 52% without error signals
- •Scanned PDFs failed text extraction checks but were marked as processed, losing 32% of corpus
- •OCR recovery and deduplication improved coverage from 68.5% to 99.6% and accuracy from 58.3% to 63.9%
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



