Added single PDF end-to-end: ingest, query, streamed answer

[markgewhite]

Summary

• Added DocumentLoader, Chunker, VectorStore, and Answerer modules implementing the first end-to-end RAG slice
• Wired Chainlit on_message handler: query → semantic search → streamed LLM answer
• LangChain used only for PDF loading (PyPDFLoader) and text splitting (RecursiveCharacterTextSplitter), with boundary comments throughout
• All other components (embeddings, vector store, prompt construction, LLM calls) use direct library calls (ollama, chromadb)

Closes #4

Test plan

• Tracer bullet test: Document → chunk → store → search → prompt (end-to-end path)
• Chunker tests: short/long docs, size limits, metadata preservation, chunk_index, doc_hash
• DocumentLoader tests: PDF loading, metadata, empty folder, recursive/non-recursive
• VectorStore tests: add, search, get_all_texts, has_document (in-memory ChromaDB)
• Answerer tests: prompt construction with source citations
• Manual: uv run chainlit run app.py with a configured document folder

🤖 Generated with Claude Code