feat: PageRank-based image summarizer

.vscode/settings.json

@@ -27,5 +27,7 @@
     "editor.codeActionsOnSave": {
       "source.organizeImports": "explicit"
     }
-  }
+  },
+  "python.testing.pytestArgs": ["packages/backend"],
+  "python.testing.pytestEnabled": true
 }

packages/backend/embedding_atlas/cli.py

@@ -227,6 +227,14 @@ def import_modules(names: list[str]):
     "neighbors_column",
     help='Column containing pre-computed nearest neighbors in format: {"ids": [n1, n2, ...], "distances": [d1, d2, ...]}. IDs should be zero-based row indices.',
 )
+@click.option(
+    "--pagerank",
+    "pagerank_column",
+    default=None,
+    is_flag=False,
+    flag_value="__compute__",
+    help="Compute PageRank scores from the neighbor graph, or specify a column containing pre-computed scores. Automatically computed when --image is specified.",
+)
 @click.option(
     "--query",
     default=None,
@@ -352,6 +360,7 @@ def main(
     x_column: str | None,
     y_column: str | None,
     neighbors_column: str | None,
+    pagerank_column: str | None,
     query: str | None,
     sample: int | None,
     umap_n_neighbors: int | None,
@@ -475,12 +484,32 @@ def main(
         labels_df = load_pandas_data(labels)
         labels_resolved = labels_df.to_dict("records")
 
+    # Compute PageRank from neighbor graph when requested or when --image is specified
+    should_compute_pagerank = (pagerank_column == "__compute__") or (
+        image is not None and pagerank_column is None
+    )
+    if (
+        should_compute_pagerank
+        and neighbors_column is not None
+        and neighbors_column in df.columns
+    ):
+        from embedding_atlas.pagerank import compute_pagerank_column
+
+        logger.info("Computing PageRank scores from neighbor graph...")
+        pagerank_column = find_column_name(df.columns, "pagerank")
+        df[pagerank_column] = compute_pagerank_column(df, neighbors=neighbors_column)
+    elif pagerank_column == "__compute__":
+        logger.warning("Cannot compute PageRank: no neighbor data available.")
+        pagerank_column = None
+
     props = make_embedding_atlas_props(
         row_id=id_column,
         x=x_column,
         y=y_column,
         neighbors=neighbors_column,
+        importance=pagerank_column,
         text=text,
+        image=image,
         point_size=point_size,
         stop_words=stop_words_resolved,
         labels=labels_resolved,

packages/backend/embedding_atlas/options.py

@@ -17,6 +17,13 @@ class EmbeddingAtlasOptions(TypedDict, total=False):
     text:
         The column name for the textual data.
 
+    image:
+        The column name for image data.
+
+    importance:
+        The column name for importance scores (e.g., PageRank). Used with ``image`` to select
+        representative images for cluster labels. Maps to ``importance`` in the frontend API.
+
     neighbors:
         The column name containing precomputed K-nearest neighbors for each point.
         Each value in the column should be a dictionary with the format:
@@ -60,6 +67,8 @@ class EmbeddingAtlasOptions(TypedDict, total=False):
     x: str | None
     y: str | None
     text: str | None
+    image: str | None
+    importance: str | None
     neighbors: str | None
 
     point_size: float | None
@@ -109,6 +118,8 @@ def set_prop(key: str, value):
     if options.get("x") is not None and options.get("y") is not None:
         set_prop("data.projection", {"x": options.get("x"), "y": options.get("y")})
     set_prop("data.text", options.get("text"))
+    set_prop("data.image", options.get("image"))
+    set_prop("data.importance", options.get("importance"))
     set_prop("data.neighbors", options.get("neighbors"))
 
     # Embedding View