rag

Prompt file imported from MikahNiehaus/ClaudeBoost (.claude/commands/rag.md). Fill in {{arguments}}, {{CLAUDEBOOST_PYTHON}}, {{CLAUDEBOOST_HOME}}, {{TEMP}} before use. Copyright stays with the author.

/rag — RAG Server Start / Reconnect

Arguments: {{arguments}}

Use this skill to start the RAG server at the beginning of a session, or reconnect after a disconnect. All verification uses the HTTP REST API — no MCP required.

Step 1: Start Both RAG Servers via Supervisor

"{{CLAUDEBOOST_PYTHON}}" "{{CLAUDEBOOST_HOME}}/scripts/rag-supervisor.py" start

This starts the supervisor managing both the main RAG server (port 8612) and clean-rag server (port 8613) with auto restart on crash. If the supervisor is already running, it reports existing status.

If the output says "already running" or shows both servers alive, proceed. If it fails, fall back to direct start:

"{{CLAUDEBOOST_PYTHON}}" "{{CLAUDEBOOST_HOME}}/scripts/rag-server-start.py"

If both fail, stop and report the error.

Step 2: Verify Both Servers via HTTP

Poll main RAG server until ready (up to 60s):

for attempt in $(seq 1 20); do
  STATUS=$(curl -s --max-time 3 http://127.0.0.1:8613/status)
  echo "$STATUS" | grep -q '"status": *"ready"' && break
  echo "waiting for main RAG server... (attempt $attempt)"
  sleep 3
done
echo "$STATUS"

Then check clean-rag server:

for attempt in $(seq 1 10); do
  CR_STATUS=$(curl -s --max-time 3 http://127.0.0.1:8613/status)
  echo "$CR_STATUS" | grep -q '"status"' && break
  echo "waiting for clean-rag server... (attempt $attempt)"
  sleep 3
done
echo "$CR_STATUS"

($attempt is a loop variable and $STATUS/$CR_STATUS are assigned in the same command, so bash-guard allows them; curl to 127.0.0.1 is fine.)

If exit code is non-zero, stop and tell the user: "RAG server did not start. Check the terminal for errors or run python $CLAUDEBOOST_HOME/scripts/rag-server-start.py manually."

Parse the main RAG JSON output and note:

• model — embedding model name
• embedding_dimensions — vector size
• collections.knowledge — chunk and file counts
• collections.agents — chunk and file counts
• indexed_projects — any project codebases currently indexed

Step 3: Write the RAG sentinel

The sentinel tells session-primer.py that RAG is verified for this session:

touch "{{TEMP}}/claudeboost_rag_ok"

(Brace form {{TEMP}} — bash-guard blocks bare $TEMP because Claude Code's expansion scanner prompts on it.)

Step 4: Prime the session

Run one real search so the embedding model is loaded and the index is proven readable:

curl -s --max-time 15 -X POST http://127.0.0.1:8613/search \
  -H "Content-Type: application/json" \
  -d '{"query":"project structure and conventions","sources":["project:'"$CLAUDEBOOST_HOME"'"],"mode":"both","limit":5}'

Read the JSON response: a results array means RAG is primed; an error field means it failed (note it in the report).

If results came back, RAG is fully live. An empty results array still counts as live — the call was served, nothing scored above the threshold. If it errors but /status was healthy, the model may still be warming up — note this in the report and the user can re-run /rag in 30s.

Step 5: Report

Success:

RAG live — main: port 8612 | knowledge: Xc/Yf agents: Xc/Yf | clean-rag: port 8613 | supervisor: auto restart active | session primed

Model still loading (status was ready but /search timed out):

RAG starting — model loading (~60s). Run /rag again when ready. Status line shows RAG indicator when live.

Failed:

RAG failed — [specific error]. Try: "{{CLAUDEBOOST_PYTHON}}" "{{CLAUDEBOOST_HOME}}/scripts/rag-supervisor.py" start in the terminal.

Notes

• To do a full session activation (GT, hooks, workspace discovery, etc.), use /boost instead.
• To reindex knowledge and agent files after changes, run /index-boost.
• The HTTP REST API docs are in knowledge/rag-http-api.xml.