scientific rag pipeline with hybrid retrieval, multi-style citations, and semantic scholar integration

[ykd007]

This adds a full scientific RAG pipeline under a rag/ module. Here's what's in it:

rag/pipeline.py — ScientificRAGPipeline
The main class that ties everything together. Accepts uploaded files or Semantic Scholar paper IDs, retrieves the most relevant chunks for a query using hybrid search, then generates a grounded answer via OpenAI. Supports both blocking and streaming (SSE) modes.

rag/hybrid_retriever.py — HybridRetriever
Combines BM25 keyword search (pure Python, no deps) with dense cosine similarity via any embedding function. The alpha parameter lets you tune how much weight to give each signal. Falls back to BM25 only if embeddings aren't available. Per-user isolation is built in so one user can't see another's private documents.

rag/document_manager.py — DocumentManager
Handles PDF (pdfminer or PyPDF2), DOCX (python-docx), TXT, MD, and HTML. Chunks with configurable size and overlap. Deduplicates by content hash so uploading the same file twice is a no-op.

rag/semantic_scholar.py — SemanticScholarClient
Thin wrapper around the Semantic Scholar Graph API. Supports search, single-paper fetch, references, and citations. Handles 429 rate limiting with automatic backoff. Works without an API key (1 req/s) or faster with one.

rag/citation_manager.py — CitationManager
Formats paper dicts into APA, MLA, Chicago, IEEE, Harvard, Vancouver, or BibTeX strings.

rag/api_routes.py — FastAPI router
Mounts under /rag — endpoints for document upload, document list/delete, paper search, paper ingest, blocking query, and SSE streaming query. Just app.include_router(rag_router) to wire it in.

tests/test_rag.py
16 unit tests, all passing. Cover BM25 retrieval, user isolation, vector blending, all 7 citation styles, document deduplication, unsupported file rejection, and mocked S2 API calls.

Closes #45