Add wizard for docker-compose creation

.gitignore

@@ -27,3 +27,6 @@ pnpm-lock.yaml
 
 .astro
 *.old
+
+# build-time data caches
+.cache/

package.json

@@ -33,6 +33,7 @@
     "astro-seo": "^1.1.0",
     "limax": "4.2.3",
     "lodash.merge": "^4.6.2",
+    "prismjs": "^1.30.0",
     "unpic": "^4.2.2"
   },
   "devDependencies": {
@@ -47,6 +48,7 @@
     "@types/js-yaml": "^4.0.9",
     "@types/lodash.merge": "^4.6.9",
     "@types/mdx": "^2.0.13",
+    "@types/prismjs": "^1.26.5",
     "@typescript-eslint/eslint-plugin": "^8.60.0",
     "@typescript-eslint/parser": "^8.60.0",
     "astro-compress": "^2.4.1",

src/assets/styles/tailwind.css

@@ -19,6 +19,9 @@
   --font-serif: 'var(--aw-font-serif, ui-serif)', ui-serif, Georgia, Cambria, 'Times New Roman', Times, serif;
   --font-heading: 'var(--aw-font-heading, ui-sans-serif)', ui-sans-serif, system-ui, sans-serif;
 
+  --text-xxs: 0.65rem;
+  --text-xxs--line-height: 1rem;
+
   --animate-fade: fadeInUp 1s both;
 
   @keyframes fadeInUp {

src/components/widgets/Stats.astro

@@ -22,8 +22,8 @@ const {
   <div class="flex flex-wrap justify-center -m-4 text-center">
     {
       stats &&
-        stats.map(({ amount, title, icon }) => (
-          <div class="p-4 md:w-1/4 sm:w-1/2 w-full min-w-[220px] text-center md:border-r md:last:border-none dark:md:border-slate-500">
+        stats.map(({ amount, title, icon, disclaimer }) => (
+          <div class="p-4 md:w-1/4 sm:w-1/2 w-full min-w-55 text-center md:border-r md:last:border-none dark:md:border-slate-500">
             {icon && (
               <div class="flex items-center justify-center mx-auto mb-4 text-primary">
                 <Icon name={icon} class="w-10 h-10" />
@@ -39,6 +39,9 @@ const {
                 {title}
               </div>
             )}
+            { disclaimer && (
+              <div class="text-xxs text-gray-600 dark:text-slate-500">{disclaimer}</div>
+            )}
           </div>
         ))
     }

src/content/docs/docs/features/facial-recognition.md

@@ -115,7 +115,7 @@ The service is configured independently, via its own `.env` file (copy `.env.exa
 | `VISION_FACE_MAX_FACES_PER_PHOTO`   | `10`    | Maximum faces included in a callback payload.                           |
 | `VISION_FACE_MIN_FACE_SIZE_PIXELS`  | `0`     | Minimum face size in pixels; `0` disables this filter.                  |
 | `VISION_FACE_BLUR_THRESHOLD`        | `0.5`   | Laplacian variance threshold; blurry faces below this are discarded.    |
-| `VISION_FACE_CLUSTER_EPS`           | `0.6`   | DBSCAN epsilon (max cosine distance) used for face clustering.          |
+| `VISION_FACE_CLUSTER_EPS`           | `0.3`   | DBSCAN epsilon (max cosine distance) used for face clustering.          |
 
 ### Storage
 
@@ -193,7 +193,7 @@ The container exposes interactive API docs at `/docs` and a health check at `/he
 - Try a different `VISION_FACE_DETECTOR_BACKEND` (`retinaface`, `mtcnn`, `opencv`, `ssd`) — detectors vary in recall depending on pose, lighting, and image resolution.
 
 **Clustering groups unrelated faces together, or fails to group similar faces:**
-- Clustering is controlled by `VISION_FACE_CLUSTER_EPS` (default `0.6`), the DBSCAN epsilon (maximum cosine distance) allowed within a cluster. Lower it if dissimilar people are being grouped together; raise it if the same person is being split across multiple clusters.
+- Clustering is controlled by `VISION_FACE_CLUSTER_EPS` (default `0.3`), the DBSCAN epsilon (maximum cosine distance) allowed within a cluster. Lower it if dissimilar people are being grouped together; raise it if the same person is being split across multiple clusters.
 - To apply a new value: update `VISION_FACE_CLUSTER_EPS` in the microservice's `.env`, restart the microservice so it picks up the change, then re-trigger clustering from _Settings ⇒ Maintenance ⇒ Run Clustering_ in Lychee. This re-clusters every currently unassigned face with the new epsilon — repeat with a different value if the result still isn't right.
 - The `VISION_FACE_MODEL_NAME` (default `ArcFace`) recognition model determines embedding quality — a different model may produce more consistent embeddings for your photo set, but changing it does not retroactively update existing embeddings.
 - Poor detections (blurry, low-confidence, or partial faces — see above) produce noisy embeddings that cluster poorly; fixing detection quality often improves clustering as a side effect.