fix(retriever): correctly cap PDF extract actors to avoid CPU oversubscription

nemo_retriever/src/nemo_retriever/ingest_modes/batch.py

@@ -1581,6 +1581,22 @@ def _build_graph_node_overrides(self) -> dict[str, dict[str, Any]]:
             embed_gpus = self._requested_plan.get_embed_gpus_per_actor()
             embed_concurrency = self._requested_plan.get_embed_initial_actors()
 
+        # Cap PDF extract actors so they don't exhaust the CPU budget.  The
+        # overhead here is the sum of CPUs that other persistent actors in the
+        # pipeline hold simultaneously: DocToPdf, PDFSplit, UDFOperator, and
+        # the ReadBinary task (4 total) plus the initial actors for
+        # page-elements, OCR, and embed.
+        non_pdf_cpu_overhead = (
+            4
+            + page_elements_concurrency * page_elements_cpus
+            + ocr_concurrency * ocr_cpus
+            + embed_concurrency * embed_cpus
+        )
+        pdf_extract_tasks = min(
+            pdf_extract_tasks,
+            max(1, int((self._total_cpu_count - non_pdf_cpu_overhead) // pdf_extract_cpus)),
+        )
+
         overrides: dict[str, dict[str, Any]] = {
             "DocToPdfConversionActor": {"batch_size": 1, "num_cpus": 1},
             "PDFSplitActor": {"batch_size": pdf_split_batch_size, "num_cpus": 1},