Mrag

mosadd Knowledge Base (mRAG)

This skill operates the mRAG OS module of mosadd — semantic recall over the user's own communications, grounded strictly in their indexed mosadd data. Hybrid vector + BM25 retrieval with Cohere reranking and a freshness lane for just-sent items.

When to invoke

Trigger on these user intents:

• "What did <contact> say about <topic>?" — mRAG_search
• "Summarize my <email/call/DM> threads about <X>" — mRAG_search
• "Find the message where I mentioned <X>" — mRAG_search
• "Ingest this document into my knowledge base" — mRAG_ingest

How to operate

1. mRAG_search is grounded, not generative. The tool returns an answer + cited source chunks (source_type, source_id, thread_id, snippet, score). Never claim something the cited sources don't support; if the answer says "no matching sources", say so plainly to the user — don't fabricate from training data.

2. Filter by source type when the user is specific. "What did I email Alice about Q3?" → source_types: ["email"]. "Recap my last call with Bob about pricing." → source_types: ["call"]. Leave source_types empty to search everything.

3. Top-K is a budget. top_k (default 8, max 25) bounds retrieval depth. For "find the one email where I mentioned X" prefer top_k: 3; for "summarize everything about <topic>" go higher.

4. Use thread_id when the user pins a single conversation. e.g. "Search just my work thread with Alice" → resolve the thread via mDM_list, then pass that thread_id to mRAG_search.

5. mRAG_ingest is for net-new content. Messages, emails, and calls already-sent are indexed automatically by the mosadd backend. Use mRAG_ingest only for external documents/notes the user wants searchable alongside their communications.

6. Cite sources back to the user. When you summarize a search result, tell the user which messages/emails/calls back the claim ("from your March 12 email to Bob" / "from yesterday's call with Alice"). The source_id + thread_id make this exact.

Configuration

mRAG needs the Supabase backend wired (it queries the same rag-query Edge Function that powers in-app RAG):

MOSADD_SUPABASE_URL=https://<your-project>.supabase.co
MOSADD_SUPABASE_ANON_KEY=<anon key>
MOSADD_USER_JWT=<your session JWT>

On the hosted hub, mRAG is gated by plan — Free has RAG off; Pro/Team have it on. BYOK / self-host has no gate; just runs against the user's own backend.

Example

User: "What did Bob say about pricing in our last call?"

You:

1. mRAG_search({ query: "Bob pricing", source_types: ["call"], top_k: 5 })
2. Result: { answer: "Bob said the SMB tier should land at $9/mo with a 14-day trial...", sources: [{ source_type: "call", source_id: "call-uuid", thread_id: "call:bob:2026-05-30", snippet: "...pricing for SMBs should be nine dollars...", score: 0.91 }] }
3. Tell the user: "From your May 30 call with Bob: he said the SMB tier should land at $9/mo with a 14-day trial."

Don't

• Don't claim mRAG results are encrypted — RAG indexes content plaintext server-side by necessity (vector search needs the embedding). The user opted in to indexing; the privacy guarantee here is "your own backend", not "zero-knowledge". See docs/security/e2ee-posture.md.
• Don't mRAG_ingest random user input as a side effect — only when the user explicitly asks to add something to their knowledge base.
• Don't use mRAG for searching the public web. mRAG is only the user's own indexed data.