Weitere Tools eingebaut. Damit ist die SQL AI einsatzbereit.

2026-01-28 22:27:34 +01:00
parent fa0a7c3b1e
commit 9b2573ebb8
5 changed files with 414 additions and 102 deletions
--- a/README.md
+++ b/README.md
@@ -1,21 +1,51 @@
-# jr-sql-ai (Terminal-first SQL Server Expert KI, lokal)
+# jr-sql-ai (Terminal-first SQL Server 2022 Expert KI, lokal)

-Ziel: Lokale Expert-KI für **SQL Server 2022** (T-SQL, Views, Stored Procedures, Execution Plans, UTF-8 Migration),
+Lokale Expert-KI für **SQL Server 2022** (T-SQL, Views, Stored Procedures, Execution Plans, UTF-8 Migration),
 aufrufbar **vom Terminal**, ohne direkte DB-Verbindung (nur Copy & Paste / Dateien).

+Die KI läuft lokal via **Ollama** in Docker und wird über ein kleines CLI (`sqlai`) genutzt.
+
+---
+
 ## Features
- **Einfacher Tech-Stack:** Docker + Ollama + Bash + curl + python
- **Host Networking:** nutzt den Host-Netzwerkstack (Routing/DNS wie Host; ideal wenn nur `br0` zuverlässig ist)
- **Auto-Updates:** Runtime + Models via `systemd --user` Timer
- **Viele Logs:** jede Ausführung schreibt detaillierte Logs unter `./logs/`
- **Freie SQL Server 2022 Q&A:** Modus `ask` für allgemeine Fragen, ohne Pipe via `--text`
+
+- **Schmaler Tech-Stack:** Docker + Ollama + Bash + curl + python
+- **Terminal-first:** `sqlai ask ...` / `sqlai analyze-tsql ...` etc.
+- **Ohne DB-Verbindung:** nur Analyse/Planung/Empfehlung anhand von Input
+- **Auto-Updates:** Runtime + Models via `scripts/update.sh` und optional `systemd --user` Timer
+- **Warmup:** nach Updates (und optional nach Bootstrap) wird ein kurzer Request gesendet (schneller “First Real Query”)
+- **Selftest:** `scripts/selftest.sh` prüft End-to-End (Docker, API, Models, echte Anfrage)
+- **Resilient:** Model-Fallback (wenn Expert-Model fehlt → Base-Model)
+- **Viele Logs:** jedes Script schreibt nachvollziehbare Logs nach `./logs/`
+
+---

 ## Voraussetzungen (Arch Linux)
- docker + docker compose
- curl
- python

-## Quickstart
+- `docker` + `docker compose`
+- `curl`
+- `python`
+- optional (für GPU): NVIDIA Container Toolkit + Compose GPU-Konfiguration
+
+---
+
+## Repository Struktur (Kurzüberblick)
+
+- `bin/sqlai` – CLI (ask + Analyse-Modi) inkl. Logging & Model-Fallback
+- `scripts/bootstrap.sh` – startet Container, pullt Modelle, baut Expert-Model, Warmup
+- `scripts/update.sh` – pullt Runtime+Modelle, rebuild Expert-Model, Warmup
+- `scripts/selftest.sh` – End-to-End Check inkl. echter Anfrage
+- `docker-compose.yml` – Ollama Service (Host networking, optional GPU)
+- `Modelfile` – Expert-Systemprompt (SQL Server 2022 Fokus)
+- `prompts/` – Prompt-Library (Copy & Paste Templates)
+- `systemd/user/` – optionaler Daily Update Timer (User Service)
+
+---
+
+# 1) Installation / Build (Bootstrap)
+
+## 1.1 Repo auspacken / klonen
+Wenn du ein tar.gz erhalten hast:
 ```bash
-cp .env.example .env
-./scripts/bootstrap.sh
+tar -xzf jr-sql-ai.tar.gz
+cd jr-sql-ai
--- a/bin/sqlai
+++ b/bin/sqlai
@@ -7,6 +7,7 @@ ENV_FILE="${ROOT}/.env"

 : "${OLLAMA_URL:=http://127.0.0.1:11434}"
 : "${EXPERT_MODEL:=jr-sql-expert}"
+: "${BASE_MODEL:=}"  # fallback

 ts() { date -Is; }

@@ -14,7 +15,6 @@ log_dir="${ROOT}/logs"
 mkdir -p "$log_dir"
 log_file="${log_dir}/sqlai-$(date -I).log"

-# Log everything (stdout+stderr)
 exec > >(tee -a "$log_file") 2>&1

 usage() {
@@ -83,7 +83,6 @@ if [[ -z "${input//[[:space:]]/}" ]]; then
  exit 4
 fi

-# Short instruction per mode (core policy is in Modelfile SYSTEM prompt)
 case "$mode" in
  ask)
    instruction=$'Beantworte die Frage als SQL Server 2022 Experte.\n\nWichtig:\n- Wenn Kontext fehlt: keine Rückfragen stellen; stattdessen Annahmen offenlegen und Optionen (A/B/C) mit Vor-/Nachteilen geben.\n- Ergebnis immer strukturiert mit: Kurzfazit, Optionen, Risiken/Checks, Nächste Schritte.\n- Wenn sinnvoll: konkrete T-SQL/DDL Snippets und Checklisten liefern.'
@@ -112,12 +111,50 @@ esac

 req_id="$(date +%Y%m%d-%H%M%S)-$$-$RANDOM"

+# --- helper: model existence via /api/tags ---
+model_exists() {
+  # args: modelname
+  local m="$1"
+  local tags
+  tags="$(curl -sS "${OLLAMA_URL}/api/tags" 2>/dev/null || true)"
+  [[ -z "$tags" ]] && return 1
+  printf '%s' "$tags" | python -c '
+import json,sys
+m=sys.argv[1]
+try:
+  obj=json.load(sys.stdin)
+except Exception:
+  sys.exit(2)
+names=set()
+for it in obj.get("models", []):
+  n=it.get("name")
+  if n: names.add(n)
+# Accept exact match, or ":latest" variants
+ok = (m in names) or (m + ":latest" in names) or (m.endswith(":latest") and m[:-7] in names)
+sys.exit(0 if ok else 1)
+' "$m" >/dev/null 2>&1
+}
+
+pick_model() {
+  # Prefer expert, fallback to base
+  if model_exists "$EXPERT_MODEL"; then
+    printf '%s' "$EXPERT_MODEL"
+    return 0
+  fi
+  if [[ -n "${BASE_MODEL:-}" ]] && model_exists "$BASE_MODEL"; then
+    echo "[$(ts)] sqlai: WARN: expert model not found in /api/tags; falling back to BASE_MODEL=${BASE_MODEL}"
+    printf '%s' "$BASE_MODEL"
+    return 0
+  fi
+  echo "[$(ts)] sqlai: ERROR: neither EXPERT_MODEL=${EXPERT_MODEL} nor BASE_MODEL=${BASE_MODEL:-<empty>} found (api/tags)."
+  return 1
+}
+
 echo "================================================================================"
 echo "[$(ts)] sqlai: REQUEST_START id=$req_id"
-echo "[$(ts)] sqlai: MODE=$mode MODEL=$EXPERT_MODEL OLLAMA_URL=$OLLAMA_URL"
+echo "[$(ts)] sqlai: MODE=$mode EXPERT_MODEL=$EXPERT_MODEL BASE_MODEL=${BASE_MODEL:-<empty>} OLLAMA_URL=$OLLAMA_URL"
 echo "[$(ts)] sqlai: INPUT_SOURCE=$input_src INPUT_BYTES=$(printf "%s" "$input" | wc -c)"

-# Delimited markup
 prompt=$(cat <<EOF
 ${instruction}

@@ -127,43 +164,51 @@ ${input}
 EOF
 )

-# Build JSON safely (no heredoc+herestring combos)
-payload="$(
-printf '%s' "$prompt" | python -c '
+payload_for_model() {
+  local model="$1"
+  printf '%s' "$prompt" | python -c '
 import json, os, sys
-model=os.environ.get("EXPERT_MODEL","jr-sql-expert")
+model=sys.argv[1]
 prompt=sys.stdin.read()
 print(json.dumps({"model": model, "prompt": prompt, "stream": False}, ensure_ascii=False))
-'
-)"
+' "$model"
+}

-resp_file="$(mktemp)"
-http_code="$(
-curl -sS -o "$resp_file" -w "%{http_code}" \
-  -X POST "${OLLAMA_URL}/api/generate" \
-  -H 'Content-Type: application/json' \
-  --data-binary "$payload" \
-  || true
-)"
+do_request() {
+  # args: model, attempt
+  local model="$1"
+  local attempt="$2"

-resp="$(cat "$resp_file")"
-rm -f "$resp_file"
+  local payload resp_file http_code resp resp_bytes extracted error_msg response_txt

-echo "[$(ts)] sqlai: HTTP_CODE=$http_code RESP_BYTES=$(printf "%s" "$resp" | wc -c) id=$req_id"
+  payload="$(payload_for_model "$model")"

-# Validate JSON
-if ! printf '%s' "$resp" | python -c 'import json,sys; json.load(sys.stdin)' >/dev/null 2>&1; then
-  echo "[$(ts)] sqlai: ERROR: response is not valid JSON id=$req_id"
-  echo "[$(ts)] sqlai: RAW_RESPONSE_BEGIN id=$req_id"
-  printf '%s\n' "$resp"
-  echo "[$(ts)] sqlai: RAW_RESPONSE_END id=$req_id"
-  echo "[$(ts)] sqlai: REQUEST_END id=$req_id status=error"
-  exit 11
-fi
+  resp_file="$(mktemp)"
+  http_code="$(
+    curl -sS -o "$resp_file" -w "%{http_code}" \
+      -X POST "${OLLAMA_URL}/api/generate" \
+      -H 'Content-Type: application/json' \
+      --data-binary "$payload" \
+      || true
+  )"

-# Extract error/response/metrics in one pass
-extracted="$(
-printf '%s' "$resp" | python -c '
+  resp="$(cat "$resp_file")"
+  rm -f "$resp_file"
+
+  resp_bytes="$(printf "%s" "$resp" | wc -c)"
+  echo "[$(ts)] sqlai: attempt=$attempt model=$model HTTP_CODE=$http_code RESP_BYTES=$resp_bytes id=$req_id"
+
+  # Validate JSON
+  if ! printf '%s' "$resp" | python -c 'import json,sys; json.load(sys.stdin)' >/dev/null 2>&1; then
+    echo "[$(ts)] sqlai: ERROR: invalid JSON response attempt=$attempt id=$req_id"
+    echo "[$(ts)] sqlai: RAW_RESPONSE_BEGIN attempt=$attempt id=$req_id"
+    printf '%s\n' "$resp" | head -n 200
+    echo "[$(ts)] sqlai: RAW_RESPONSE_END attempt=$attempt id=$req_id"
+    return 90
+  fi
+
+  extracted="$(
+    printf '%s' "$resp" | python -c '
 import json,sys
 obj=json.load(sys.stdin)
 out={
@@ -180,47 +225,72 @@ out={
 }
 print(json.dumps(out, ensure_ascii=False))
 '
-)"
+  )"

-error_msg="$(printf '%s' "$extracted" | python -c 'import json,sys; print((json.load(sys.stdin).get("error") or "").strip())')"
-response_txt="$(printf '%s' "$extracted" | python -c 'import json,sys; print(json.load(sys.stdin).get("response") or "")')"
+  error_msg="$(printf '%s' "$extracted" | python -c 'import json,sys; print((json.load(sys.stdin).get("error") or "").strip())')"
+  response_txt="$(printf '%s' "$extracted" | python -c 'import json,sys; print(json.load(sys.stdin).get("response") or "")')"

-# HTTP != 200 is error
-if [[ "$http_code" != "200" ]]; then
-  echo "[$(ts)] sqlai: ERROR: non-200 HTTP_CODE=$http_code id=$req_id"
-  if [[ -n "$error_msg" ]]; then
-    echo "[$(ts)] sqlai: OLLAMA_ERROR=$error_msg id=$req_id"
-  else
-    echo "[$(ts)] sqlai: BODY_SNIPPET_BEGIN id=$req_id"
-    printf '%s\n' "$resp" | head -n 120
-    echo "[$(ts)] sqlai: BODY_SNIPPET_END id=$req_id"
+  # HTTP != 200 -> error
+  if [[ "$http_code" != "200" ]]; then
+    echo "[$(ts)] sqlai: ERROR: non-200 HTTP_CODE=$http_code attempt=$attempt id=$req_id"
+    [[ -n "$error_msg" ]] && echo "[$(ts)] sqlai: OLLAMA_ERROR=$error_msg attempt=$attempt id=$req_id"
+    return 91
+  fi
+
+  # API-level error
+  if [[ -n "$error_msg" ]]; then
+    echo "[$(ts)] sqlai: ERROR: OLLAMA_ERROR=$error_msg attempt=$attempt id=$req_id"
+    # special code for retry decisions
+    if printf '%s' "$error_msg" | grep -qiE 'model .*not found|not found'; then
+      return 42
+    fi
+    return 92
+  fi
+
+  # Print answer
+  printf "\n%s\n\n" "$(printf "%s" "$response_txt" | sed 's/[[:space:]]*$//')"
+
+  # Print metrics optionally
+  if [[ "$no_metrics" != "1" ]]; then
+    echo "METRICS=$(printf '%s' "$extracted" | python -c 'import json,sys; import json as j; print(j.dumps(json.load(sys.stdin)["metrics"], ensure_ascii=False))')"
+  fi
+
+  # Empty answer warning
+  if [[ -z "${response_txt//[[:space:]]/}" ]]; then
+    echo "[$(ts)] sqlai: WARN: empty response attempt=$attempt model=$model id=$req_id"
+  fi
+
+  return 0
+}
+
+model_primary="$(pick_model)"
+echo "[$(ts)] sqlai: selected_model_primary=$model_primary id=$req_id"
+
+# Attempt 1
+set +e
+do_request "$model_primary" "1"
+rc=$?
+set -e
+
+# Retry with BASE_MODEL if "model not found" and we weren't already using it
+if [[ "$rc" -eq 42 ]]; then
+  if [[ -n "${BASE_MODEL:-}" ]] && [[ "$model_primary" != "$BASE_MODEL" ]] && model_exists "$BASE_MODEL"; then
+    echo "[$(ts)] sqlai: WARN: retrying with BASE_MODEL=$BASE_MODEL (expert model not found) id=$req_id"
+    set +e
+    do_request "$BASE_MODEL" "2"
+    rc2=$?
+    set -e
+    rc="$rc2"
+  else
+    echo "[$(ts)] sqlai: ERROR: retry requested but BASE_MODEL unavailable id=$req_id"
+    rc=93
  fi
-  echo "[$(ts)] sqlai: REQUEST_END id=$req_id status=error"
-  exit 12
 fi

-# API-level error
-if [[ -n "$error_msg" ]]; then
-  echo "[$(ts)] sqlai: ERROR: OLLAMA_ERROR=$error_msg id=$req_id"
-  echo "[$(ts)] sqlai: REQUEST_END id=$req_id status=error"
-  exit 13
-fi
-
-# Print answer
-printf "\n%s\n\n" "$(printf "%s" "$response_txt" | sed 's/[[:space:]]*$//')"
-
-# If response empty, dump JSON snippet to log
-if [[ -z "${response_txt//[[:space:]]/}" ]]; then
-  echo "[$(ts)] sqlai: WARN: empty response id=$req_id"
-  echo "[$(ts)] sqlai: RAW_JSON_SNIPPET_BEGIN id=$req_id"
-  printf '%s\n' "$resp" | head -n 200
-  echo "[$(ts)] sqlai: RAW_JSON_SNIPPET_END id=$req_id"
-fi
-
-# Print metrics optionally
-if [[ "$no_metrics" != "1" ]]; then
-  metrics_line="$(printf '%s' "$extracted" | python -c 'import json,sys; print("METRICS="+json.dumps(json.load(sys.stdin)["metrics"], ensure_ascii=False))')"
-  echo "$metrics_line"
+if [[ "$rc" -ne 0 ]]; then
+  echo "[$(ts)] sqlai: REQUEST_END id=$req_id status=error rc=$rc"
+  echo "================================================================================"
+  exit "$rc"
 fi

 echo "[$(ts)] sqlai: REQUEST_END id=$req_id status=ok"
--- a/scripts/bootstrap.sh
+++ b/scripts/bootstrap.sh
@@ -2,11 +2,7 @@
 set -euo pipefail

 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
-
-# Create .env if missing (do not overwrite)
 cp -n "${ROOT}/.env.example" "${ROOT}/.env" || true
-
-# shellcheck disable=SC1090
 source "${ROOT}/.env"

 ts(){ date -Is; }
@@ -14,8 +10,6 @@ ts(){ date -Is; }
 log_dir="${ROOT}/logs"
 mkdir -p "$log_dir"
 log_file="${log_dir}/bootstrap-$(date -Iseconds).log"
-
-# log everything (stdout+stderr)
 exec > >(tee -a "$log_file") 2>&1

 echo "[$(ts)] bootstrap: starting (ROOT=$ROOT)"
@@ -51,7 +45,6 @@ echo "[$(ts)] bootstrap: building expert model: ${EXPERT_MODEL}"
 tmp="$(mktemp)"
 sed "s/\${BASE_MODEL}/${BASE_MODEL}/g" "${ROOT}/Modelfile" > "$tmp"

-# Copy Modelfile into container and build from explicit path (robust)
 docker cp "$tmp" ollama:/tmp/Modelfile.jr-sql-expert
 docker exec -it ollama ollama create "${EXPERT_MODEL}" -f /tmp/Modelfile.jr-sql-expert

@@ -61,15 +54,15 @@ echo "[$(ts)] bootstrap: verifying model exists..."
 docker exec -it ollama ollama list | grep -F "${EXPERT_MODEL}" >/dev/null && \
  echo "[$(ts)] bootstrap: OK: ${EXPERT_MODEL} is available."

-# End-to-end test
-if [[ ! -x "${ROOT}/bin/sqlai" ]]; then
-  echo "[$(ts)] bootstrap: ERROR: ${ROOT}/bin/sqlai not found or not executable"
-  ls -la "${ROOT}/bin" || true
-  exit 2
+# Warmup
+if [[ -x "${ROOT}/bin/sqlai" ]]; then
+  echo "[$(ts)] bootstrap: warmup: sending a short request (no-metrics)..."
+  "${ROOT}/bin/sqlai" ask --text "Warmup: reply with exactly 'OK'." --no-metrics || {
+    echo "[$(ts)] bootstrap: WARN: warmup failed (continuing)."
+  }
+  echo "[$(ts)] bootstrap: warmup: done"
+else
+  echo "[$(ts)] bootstrap: WARN: ${ROOT}/bin/sqlai not executable; skipping warmup."
 fi

-echo "[$(ts)] bootstrap: test (running one request)..."
-echo "SELECT 1;" | "${ROOT}/bin/sqlai" analyze-tsql
-echo "[$(ts)] bootstrap: test done"
-
 echo "[$(ts)] bootstrap: done"
--- a/scripts/selftest.sh
+++ b/scripts/selftest.sh
@@ -0,0 +1,212 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+ENV_FILE="${ROOT}/.env"
+
+ts(){ date -Is; }
+
+log_dir="${ROOT}/logs"
+mkdir -p "$log_dir"
+log_file="${log_dir}/selftest-$(date -Iseconds).log"
+exec > >(tee -a "$log_file") 2>&1
+
+echo "================================================================================"
+echo "[$(ts)] selftest: START ROOT=$ROOT"
+
+# -------- helpers --------
+fail() {
+  local msg="$1"
+  local code="${2:-1}"
+  echo "[$(ts)] selftest: FAIL code=$code msg=$msg"
+  echo "[$(ts)] selftest: END status=FAIL"
+  echo "================================================================================"
+  exit "$code"
+}
+
+warn() {
+  echo "[$(ts)] selftest: WARN $*"
+}
+
+ok() {
+  echo "[$(ts)] selftest: OK $*"
+}
+
+need_cmd() {
+  command -v "$1" >/dev/null 2>&1 || fail "missing command: $1" 10
+}
+
+# robust JSON check
+is_json() {
+  # reads stdin
+  python -c 'import json,sys; json.load(sys.stdin)' >/dev/null 2>&1
+}
+
+# Extract model names from /api/tags JSON and check membership
+model_in_tags() {
+  # args: modelname, tags_json_string
+  local model="$1"
+  local tags="$2"
+  printf '%s' "$tags" | python -c '
+import json,sys
+m=sys.argv[1]
+obj=json.load(sys.stdin)
+names=set()
+for it in obj.get("models", []):
+  n=it.get("name")
+  if n: names.add(n)
+ok = (m in names) or (m + ":latest" in names) or (m.endswith(":latest") and m[:-7] in names)
+sys.exit(0 if ok else 1)
+' "$model" >/dev/null 2>&1
+}
+
+# -------- preflight --------
+need_cmd docker
+need_cmd curl
+need_cmd python
+need_cmd grep
+need_cmd sed
+
+if [[ ! -f "$ENV_FILE" ]]; then
+  warn ".env not found, creating from .env.example"
+  cp -n "${ROOT}/.env.example" "${ROOT}/.env" || true
+fi
+
+# shellcheck disable=SC1090
+source "${ROOT}/.env"
+
+: "${OLLAMA_URL:=http://127.0.0.1:11434}"
+: "${EXPERT_MODEL:=jr-sql-expert}"
+: "${BASE_MODEL:=}"
+
+ok "loaded env: OLLAMA_URL=$OLLAMA_URL EXPERT_MODEL=$EXPERT_MODEL BASE_MODEL=${BASE_MODEL:-<empty>}"
+
+if [[ ! -x "${ROOT}/bin/sqlai" ]]; then
+  fail "bin/sqlai missing or not executable at ${ROOT}/bin/sqlai" 11
+fi
+
+# docker compose availability (plugin or legacy)
+if docker compose version >/dev/null 2>&1; then
+  ok "docker compose available"
+else
+  fail "docker compose not available (install docker compose plugin)" 12
+fi
+
+# container state
+if docker ps --format '{{.Names}}' | grep -qx 'ollama'; then
+  ok "container 'ollama' is running"
+else
+  warn "container 'ollama' not running; attempting 'docker compose up -d'"
+  docker compose -f "${ROOT}/docker-compose.yml" up -d || fail "docker compose up failed" 13
+  if ! docker ps --format '{{.Names}}' | grep -qx 'ollama'; then
+    fail "container 'ollama' still not running after compose up" 14
+  fi
+  ok "container 'ollama' is running after compose up"
+fi
+
+# API wait loop
+ok "waiting for Ollama API at $OLLAMA_URL ..."
+api_ok="0"
+for i in {1..120}; do
+  if curl -sS "${OLLAMA_URL}/api/tags" >/dev/null 2>&1; then
+    api_ok="1"
+    break
+  fi
+  sleep 1
+done
+[[ "$api_ok" == "1" ]] || fail "Ollama API not reachable at $OLLAMA_URL after waiting" 20
+ok "Ollama API reachable"
+
+# Fetch /api/tags and validate JSON
+tags="$(curl -sS "${OLLAMA_URL}/api/tags" || true)"
+[[ -n "$tags" ]] || fail "/api/tags returned empty body" 21
+
+if ! printf '%s' "$tags" | is_json; then
+  echo "[$(ts)] selftest: /api/tags RAW_BEGIN"
+  printf '%s\n' "$tags" | head -n 200
+  echo "[$(ts)] selftest: /api/tags RAW_END"
+  fail "/api/tags did not return JSON" 22
+fi
+ok "/api/tags returned valid JSON"
+
+# Model availability checks
+expert_present="0"
+base_present="0"
+
+if model_in_tags "$EXPERT_MODEL" "$tags"; then
+  expert_present="1"
+  ok "EXPERT_MODEL present: $EXPERT_MODEL"
+else
+  warn "EXPERT_MODEL not present in tags: $EXPERT_MODEL"
+fi
+
+if [[ -n "${BASE_MODEL:-}" ]] && model_in_tags "$BASE_MODEL" "$tags"; then
+  base_present="1"
+  ok "BASE_MODEL present: $BASE_MODEL"
+else
+  if [[ -n "${BASE_MODEL:-}" ]]; then
+    warn "BASE_MODEL not present in tags: $BASE_MODEL"
+  else
+    warn "BASE_MODEL empty in .env"
+  fi
+fi
+
+if [[ "$expert_present" != "1" && "$base_present" != "1" ]]; then
+  warn "No expert/base model found. Attempting to pull BASE_MODEL if set..."
+  if [[ -n "${BASE_MODEL:-}" ]]; then
+    docker exec -it ollama ollama pull "${BASE_MODEL}" || warn "ollama pull BASE_MODEL failed"
+    tags2="$(curl -sS "${OLLAMA_URL}/api/tags" || true)"
+    if printf '%s' "$tags2" | is_json && model_in_tags "$BASE_MODEL" "$tags2"; then
+      ok "BASE_MODEL now present after pull: $BASE_MODEL"
+      base_present="1"
+    else
+      fail "Neither EXPERT_MODEL nor BASE_MODEL available after attempted pull" 23
+    fi
+  else
+    fail "Neither EXPERT_MODEL nor BASE_MODEL available and BASE_MODEL is empty" 24
+  fi
+fi
+
+# Warmup request
+ok "warmup request via sqlai (no-metrics)"
+set +e
+"${ROOT}/bin/sqlai" ask --text "Warmup: reply with exactly 'OK'." --no-metrics
+rc=$?
+set -e
+if [[ "$rc" -ne 0 ]]; then
+  warn "warmup failed rc=$rc (continuing to real query)"
+else
+  ok "warmup succeeded"
+fi
+
+# Real query (slightly longer)
+ok "real query via sqlai (ask)"
+set +e
+out="$("${ROOT}/bin/sqlai" ask --text "Give a concise checklist to troubleshoot parameter sniffing in SQL Server 2022. Keep it technical." --no-metrics 2>&1)"
+rc=$?
+set -e
+if [[ "$rc" -ne 0 ]]; then
+  echo "[$(ts)] selftest: sqlai output BEGIN"
+  printf '%s\n' "$out" | tail -n 200
+  echo "[$(ts)] selftest: sqlai output END"
+  fail "real query failed rc=$rc" 30
+fi
+
+# Basic sanity: must contain at least some non-empty text
+if [[ -z "${out//[[:space:]]/}" ]]; then
+  fail "real query returned empty output" 31
+fi
+
+ok "real query returned non-empty output"
+
+# Summary
+echo "--------------------------------------------------------------------------------"
+echo "[$(ts)] selftest: SUMMARY"
+echo "[$(ts)] selftest: API=OK"
+echo "[$(ts)] selftest: EXPERT_MODEL_PRESENT=$expert_present"
+echo "[$(ts)] selftest: BASE_MODEL_PRESENT=$base_present"
+echo "[$(ts)] selftest: LOG_FILE=$log_file"
+echo "--------------------------------------------------------------------------------"
+
+echo "[$(ts)] selftest: END status=OK"
+echo "================================================================================"
--- a/scripts/update.sh
+++ b/scripts/update.sh
@@ -2,8 +2,6 @@
 set -euo pipefail

 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
-
-# shellcheck disable=SC1090
 source "${ROOT}/.env"

 ts(){ date -Is; }
@@ -11,8 +9,6 @@ ts(){ date -Is; }
 log_dir="${ROOT}/logs"
 mkdir -p "$log_dir"
 log_file="${log_dir}/update-$(date -Iseconds).log"
-
-# log everything (stdout+stderr)
 exec > >(tee -a "$log_file") 2>&1

 echo "[$(ts)] update: starting (ROOT=$ROOT)"
@@ -60,4 +56,15 @@ echo "[$(ts)] update: verifying model exists..."
 docker exec -it ollama ollama list | grep -F "${EXPERT_MODEL}" >/dev/null && \
  echo "[$(ts)] update: OK: ${EXPERT_MODEL} is available."

+# Warmup (loads model + GPU kernels, reduces first real query latency)
+if [[ -x "${ROOT}/bin/sqlai" ]]; then
+  echo "[$(ts)] update: warmup: sending a short request (no-metrics)..."
+  "${ROOT}/bin/sqlai" ask --text "Warmup: reply with exactly 'OK'." --no-metrics || {
+    echo "[$(ts)] update: WARN: warmup failed (continuing)."
+  }
+  echo "[$(ts)] update: warmup: done"
+else
+  echo "[$(ts)] update: WARN: ${ROOT}/bin/sqlai not executable; skipping warmup."
+fi
+
 echo "[$(ts)] update: complete"