Language IntelligencePrototypeCAS Research · Internal Engineering Demonstration
AI Knowledge System
An internal demonstration of grounded answering over a curated technical corpus, with layered grounding metrics and citation precision tracking.
Findings so far
- Evidence recall@8 predicted end-to-end answer quality better than any generation-side change.
- Chunking strategy shifts moved citation precision more than model upgrades did.
- A small reranker between grounding and generation paid for its latency in quality.
Stack
EmbeddingsVector searchRerankingDecoder-only LLM