Conversation / RAG

Unlike the other connectors on this page, the conversation module doesn't pull data in — it gives Jarvis a local semantic layer over everything already captured. It embeds your Obsidian vault with fastembed (no cloud call for embedding), stores the vectors in a sidecar sqlite-vec database, and answers questions through an agentic RAG loop, an MCP server, and an in-app chat tab. The only cloud hop in the whole module is the redacted DeepSeek call inside the answer loop itself.

Env vars

Env var	Purpose
JARVIS_VECTOR_DB	Sidecar vector-store path (default: vectors.db under $JARVIS_PB_DATA or pb_data)
JARVIS_EMBED_MODEL	Embedding model id (default a multilingual MiniLM model)
JARVIS_INDEX_EXCLUDE	Extra :-separated exclude globs when indexing
JARVIS_VAULT_DIR	Vault root to index, shared with the capture pipeline
DEEPSEEK_API_KEY	Needed for the answer loop's DeepSeek hops (see LLM providers)

Where to go next

This page is just the connector-index entry — the real depth lives here:

• Semantic index — building and rebuilding the vault embedding index that everything else queries.
• Chat & RAG — asking Jarvis questions, in the chat tab or otherwise, and how citations work.
• MCP server — the tool surface (ask, search_vault, query_records, person_timeline, list_inbox) exposed to Claude Code and other MCP clients.