NovelVerse

Reading fiction online is still a solitary, linear act. Readers cannot interrogate lore, trace character arcs, or ask why a name keeps appearing. Authors managing collaborative translations juggle files and email threads. Neither side has real tooling.

Spoiler safety was the core design constraint. A naive RAG system retrieving from the full corpus would leak plot. The solution: a position-aware retrieval policy that locks semantic search to the reader's confirmed chapter boundary. Neo4j handles entity relationships and character graph traversal; pgvector handles semantic similarity. LangGraph orchestrates the compound query, choosing graph traversal versus embedding search per question type. The narration engine follows the same principle: expression is inferred, not scripted. An instruction LLM annotates each passage with inline emotion tags before synthesis, so a chase and a quiet farewell sound different without the author scripting either. Authors define a voice identity for their story; readers can override it.