Command

PePersonal JEJust Electrolyte RTRelease Timer TtTacTile CSClearSource AtAthletic

TacTile›Launch TacTile›RAG / Semantic Search·16 in this wave

Approve: RAG / Semantic SearchTo do · blocker Decide: RAG stays internal-only or becomes the agent's retrieval layerTo do · blocker Decide: will RAG ever index per-account/customer content?To do · blocker Verify RAG_QUERY_STUB is removedDone Build rag.search MCP tool and wire it into the customer agent turnTo do Fix score-floor fallback that silently answers from irrelevant chunksTo do Build a retrieval-quality eval harnessTo do Fix dead repos.yml config and scope the indexer's whole-repo catch-allTo do Add unit test coverage for chunkers, rerank, and collector logicTo do Run e2e retrieval smoke testTo do Confirm vector DB has dataTo do Add rate limiting to /api/rag/queryDone Verify RAG honest-degrade behavior after the score-floor fixTo do Run the eval harness and record a retrieval-quality baselineTo do Fix full-repo re-embed on every git push (unbounded per-push cost)To do Bound and monitor total OpenAI spend for RAGTo do

Wave

RAG / Semantic Search

Retrieval for the flagship agent — protected alongside it, not cut. Earlier framing as cut-list item #1 was wrong and is reversed. Critical finding: RAG is NOT actually wired to the customer agent today — searchAndAnswer() is only reachable via an internal endpoint behind a static token, and the corpus is TacTile's own monorepo and business docs rather than customer data. So it is currently an internal dev tool. Tenant isolation is safe only by accident: there is no account_id column and no scoping primitive, so the day customer content is added, cross-account leakage needs zero code changes to happen.

TacTile › Launch TacTile › RAG / Semantic Search

2/16 done14 open3 open blockers

Selected taskclear

Decide: will RAG ever index per-account/customer content?

To doOwner · met_55f88547Blocker

Unlocks 2Approve: RAG / Semantic Search Build rag.search MCP tool and wire it into the customer agent turn

Show the prompt

You are Viv, the agent for the TacTile lane of Command.

Open the lane and work there:

```bash
cd ~/Command/tactile && claude
```

Read that lane's CLAUDE.md and lane.json before you touch anything — they
carry your identity, your scope, and the hazards specific to this lane.
Read state from the hub (https://internal.tactile.to) or ask. Never guess at it.

## The task

**Decide: will RAG ever index per-account/customer content?** · to do · blocker

Where it sits: TacTile › Launch TacTile › RAG / Semantic Search

## Unlocks 2

- Approve: RAG / Semantic Search
- Build rag.search MCP tool and wire it into the customer agent turn

## The brief on file

me — tenant-isolation decision. Audited and confirmed SAFE TODAY: rag_document/rag_chunk/rag_embedding have RLS enabled with explicit deny-all policies for anon/authenticated (migrations 20260417164801_auth_rls_hardening_phase_1.sql + 20260504_rls_server_only.sql — only service_role/server code can reach them), the corpus contains zero customer data (it's TacTile's own git repos + iCloud company docs, ingested via packages/rag-core/indexer/ingest.ts and ingest-docs.ts), and RAG isn't reachable by the customer-facing agent at all right now. BUT there is no tenant-scoping primitive anywhere: rag_document has no owner/account_id column, and vectorSearch() (packages/rag-core/retrieval/retrieve.ts) runs with zero WHERE-clause filtering — it is architecturally impossible to scope a query to one account today. If RAG ever gets wired into the customer agent (see the internal-vs-agent decision task) and anyone later adds per-account content (e.g. a customer's own notes/destinations for personalized answers) without designing this first, cross-account leakage becomes trivial. Done when: decision recorded — corpus stays TacTile-only forever (document this constraint loudly in AGENTS.md), or a scoping design (account_id column + enforced WHERE filter) is commissioned before any customer content is ever indexed. File: packages/rag-core/retrieval/retrieve.ts, supabase/migrations/ (rag_document schema).

## How to work this

- Do the work, don't just plan it. If the fix is obvious and safe, make it.
- Surface anything needing Matthew's sign-off instead of deciding it yourself.
- Short replies. Lead with the answer, then stop.
- Finish by saying what changed, what you verified, and what you'd do next.

To doIn progressDoneBlockedPick a wave on the left · tap a task to expand it

Lanes

RAG / Semantic Search

Decide: will RAG ever index per-account/customer content?