feat: enhanced scientific RAG pipeline with cross-encoder reranking and metadata filtering

[hobyt-aluzar]

Enhanced Scientific RAG Pipeline with Cross-Encoder Reranking

This PR introduces an enhanced scientific RAG pipeline that significantly improves retrieval quality and citation accuracy, enabling fully private, local semantic reranking.

Key Changes:

• Local Reranking: Integrated @xenova/transformers with Xenova/ms-marco-MiniLM-L-6-v2 cross-encoder. It fetches an expanded pool of documents from ChromaDB, scores them semantically against the user's query, and returns the top results. This process runs 100% offline.
• Scientific Chunking & Metadata: Enhanced inject-documents.ts to use targeted text separators for scientific sections (e.g., Abstract, Methods, Results). It now extracts the section, page, and dynamically generated fallback titles to build robust metadata.
• Citation-Aware Generation: Modified rag-chat.ts to pass the Rerank Score and precise citation keys (e.g., [document-name:p1:c2]) directly into the LLM context. The prompt has been updated to enforce the usage of these citation keys, guaranteeing traceable and accurate responses.
• Test Coverage: Added scientific-rag.test.ts to ensure stability of citation keys, title normalization, and section detection.

This architecture scales cleanly and keeps all data on the user's local machine, while delivering near state-of-the-art semantic retrieval performance.