diff --git a/src/embeddings/embeddings_servable.cpp b/src/embeddings/embeddings_servable.cpp
index 7adcba6a8e..714b767368 100644
--- a/src/embeddings/embeddings_servable.cpp
+++ b/src/embeddings/embeddings_servable.cpp
@@ -415,11 +415,14 @@ std::shared_ptr<ov::Model> EmbeddingsServable::applyPrePostProcessing(ov::Core&
         postProcInferRequestsQueue = std::make_unique<OVInferRequestsQueue>(postProcCompiledModel, numberOfParallelInferRequests);
         npuPostprocessingRequired = true;
 
-        // These are the settings for NPU model
-        if (getMaxModelLength().has_value()) {
-            config.max_length = getMaxModelLength().value();
-        }
-        // Models other than Qwen requires reshaping to static shape to work on NPU.
+        // NPU model settings.
+        //
+        // For short-context models (max_position_embeddings < 1024) we reshape to a static
+        // shape and pin config.max_length to the model's max_position_embeddings, since the
+        // NPU compiler needs a static sequence length for those.
+        //
+        // For long-context / Qwen-style dynamic models we deliberately do NOT propagate
+        // max_position_embeddings into config.max_length.
         if (getMaxModelLength().has_value() && getMaxModelLength().value() < 1024) {
             modelIsStatic = true;
             config.padding_side = "right";
diff --git a/src/rerank/BUILD b/src/rerank/BUILD
index b703aa1117..4db374994a 100644
--- a/src/rerank/BUILD
+++ b/src/rerank/BUILD
@@ -20,6 +20,7 @@ load("//:common_settings.bzl", "ovms_cc_library")
 ovms_cc_library(
     name = "rerank_servable",
     hdrs = ["rerank_servable.hpp"],
+    srcs = ["rerank_servable.cpp"],
     deps = ["//src:sidepacket_servable",
             "//src/port:rapidjson_document",
             "//src/port:rapidjson_istreamwrapper",
diff --git a/src/rerank/rerank_servable.hpp b/src/rerank/rerank_servable.hpp
index f287d5b60f..26cd25fda2 100644
--- a/src/rerank/rerank_servable.hpp
+++ b/src/rerank/rerank_servable.hpp
@@ -50,6 +50,8 @@ struct RerankServable : SidepacketServable {
             addBosToken = false;
         }
     }
+
+    std::shared_ptr<ov::Model> applyPrePostProcessing(ov::Core& core, std::shared_ptr<ov::Model> model, ov::AnyMap& properties) override;
 };
 
 using RerankServableMap = std::unordered_map<std::string, std::shared_ptr<RerankServable>>;