feat: add BM25 hybrid search with RRF, update chunking for Notion ing…

[RubyRyn]

• Add bm25s dependency and BM25Manager (build/search/save/load with page title indexing)
• Add HybridRetriever with RRF merge (k=60): vector(15) + BM25(15) -> top 15
• Replace direct ChromaDB query with HybridRetriever in both conversation endpoints
• Bump reranker top_k from 3 to 10, context cap from 4000 to 5000 chars
• Remove MarkdownHeaderTextSplitter from Notion ingestion; use RecursiveCharacterTextSplitter only
• Increase Notion chunk_size from 500 to 1000, overlap from 100 to 200
• Build and save BM25 index after every Notion ingestion run

[emmanuelkb]

@RubyRyn Conversation history removed — Both send_message and stream_message no longer pass conversation_history to gemini.ask_workmate(). This means the LLM loses multi-turn
context within a conversation. Was this intentional? If ask_workmate still accepts that parameter, this is a regression

[emmanuelkb]

New env var VOYAGE_API_KEY required — No .env.example update, no documentation, and the app will crash at startup if the reranker is initialized without it. Should be
documented and ideally fail gracefully or lazily.

[RubyRyn]

• Fixed Conversation history missing issue.
• .env.example - Added GIMINI_API_KEY, VOYAGE_API_KEY and NOTION_TOKEN.
• config.py - Added VOYAGE_API_KEY: str = "" so it's tracked through Pydantic settings like other keys.
• voyage_reranker.py - No longer raises on missing key. Sets self.client = None with a waring log, and rerank() returns chunks[:top_k] unranked when the client is absent. The voyageai import is also deferred to only when the key is present.

[emmanuelkb]

great