diff --git a/core/src/features/rag/rac_rag_proto_abi.cpp b/core/src/features/rag/rac_rag_proto_abi.cpp
index 018291f68a..ed0a7e8ebd 100644
--- a/core/src/features/rag/rac_rag_proto_abi.cpp
+++ b/core/src/features/rag/rac_rag_proto_abi.cpp
@@ -422,11 +422,18 @@ rac_result_t execute_rag_query(const std::shared_ptr<Session>& s,
     // Base off RAC_LLM_OPTIONS_DEFAULT, then apply the embedded generation
     // knobs with the RAG pipeline defaults (max 512 tokens, top_p 0.9) when
     // unset.
+    //
+    // temperature and top_k are `optional` in llm_options.proto, so presence
+    // is what "unset" means: reading the value instead yields 0, which is a
+    // legal explicit setting for both (greedy / top-k disabled) and therefore
+    // cannot double as a sentinel. has_generation() is not a substitute -- a
+    // caller that sets any other generation knob makes the submessage present
+    // while leaving these two absent.
     rac_llm_options_t opts = RAC_LLM_OPTIONS_DEFAULT;
     opts.max_tokens = gen.max_output_tokens() > 0 ? gen.max_output_tokens() : 512;
-    opts.temperature = query_proto.has_generation() ? gen.temperature() : opts.temperature;
+    opts.temperature = gen.has_temperature() ? gen.temperature() : opts.temperature;
     opts.top_p = gen.top_p() > 0.0f ? gen.top_p() : 0.9f;
-    opts.top_k = gen.top_k();
+    opts.top_k = gen.has_top_k() ? gen.top_k() : opts.top_k;
     opts.disable_thinking =
         (gen.has_reasoning() && gen.reasoning().mode() == runanywhere::v1::REASONING_MODE_OFF)
             ? RAC_TRUE
diff --git a/core/tests/test_advanced_modality_proto_abi.cpp b/core/tests/test_advanced_modality_proto_abi.cpp
index bb1948d59c..26c84bf00a 100644
--- a/core/tests/test_advanced_modality_proto_abi.cpp
+++ b/core/tests/test_advanced_modality_proto_abi.cpp
@@ -103,6 +103,7 @@ std::atomic<bool> g_dummy_embeddings_started{false};
 std::atomic<bool> g_dummy_embeddings_release{false};
 float g_dummy_llm_last_temperature = -1.0f;
 int32_t g_dummy_llm_last_max_tokens = 0;
+int32_t g_dummy_llm_last_top_k = -1;
 rac_bool_t g_dummy_llm_last_disable_thinking = RAC_FALSE;
 std::string g_dummy_llm_stream_response;
 
@@ -380,6 +381,7 @@ rac_result_t dummy_llm_stream(void* impl, const char*, const rac_llm_options_t*
     if (options) {
         g_dummy_llm_last_temperature = options->temperature;
         g_dummy_llm_last_max_tokens = options->max_tokens;
+        g_dummy_llm_last_top_k = options->top_k;
         g_dummy_llm_last_disable_thinking = options->disable_thinking;
     }
     if (g_dummy_llm_block_stream.load(std::memory_order_acquire)) {
@@ -1112,6 +1114,51 @@ int test_rag_ingest_query_mocked_path() {
     CHECK(poll_capability(runanywhere::v1::CAPABILITY_OPERATION_EVENT_KIND_RAG_QUERY_COMPLETED),
           "RAG query publishes RAG_QUERY_COMPLETED");
 
+    // temperature and top_k are `optional` in llm_options.proto, so a
+    // generation submessage that sets neither must leave the
+    // RAC_LLM_OPTIONS_DEFAULT values in place rather than collapse to the
+    // proto3 zero -- 0.0f and 0 are both legal explicit settings here
+    // (greedy sampling / top-k disabled), so neither can act as a sentinel.
+    runanywhere::v1::RAGQueryOptions unset_sampling_query;
+    unset_sampling_query.set_query("Where does RAG live?");
+    unset_sampling_query.mutable_generation()->set_max_output_tokens(32);
+    std::vector<uint8_t> unset_sampling_bytes;
+    CHECK(serialize(unset_sampling_query, &unset_sampling_bytes),
+          "RAGQueryOptions with no sampling knobs serializes");
+    g_dummy_llm_last_temperature = -1.0f;
+    g_dummy_llm_last_top_k = -1;
+    rac_proto_buffer_init(&out);
+    rc = rac_rag_query_proto(session, unset_sampling_bytes.data(), unset_sampling_bytes.size(),
+                             &out);
+    CHECK(rc == RAC_SUCCESS, "RAG query with no sampling knobs succeeds");
+    CHECK(g_dummy_llm_last_temperature == RAC_LLM_OPTIONS_DEFAULT.temperature,
+          "RAG query keeps the default temperature when the field is unset");
+    CHECK(g_dummy_llm_last_top_k == RAC_LLM_OPTIONS_DEFAULT.top_k,
+          "RAG query keeps the default top_k when the field is unset");
+    rac_proto_buffer_free(&out);
+    rac_sdk_event_clear_queue();
+
+    // The other half of the same contract: 0 is a legal explicit top_k
+    // (top-k filtering disabled), so it has to survive rather than be read as
+    // "unset". This is what stops the bug from being "fixed" with a
+    // value-based sentinel such as `gen.top_k() > 0 ? gen.top_k() : default`,
+    // which would satisfy the unset case above while silently discarding a
+    // caller's explicit 0.
+    runanywhere::v1::RAGQueryOptions explicit_zero_query;
+    explicit_zero_query.set_query("Where does RAG live?");
+    explicit_zero_query.mutable_generation()->set_max_output_tokens(32);
+    explicit_zero_query.mutable_generation()->set_top_k(0);
+    std::vector<uint8_t> explicit_zero_bytes;
+    CHECK(serialize(explicit_zero_query, &explicit_zero_bytes),
+          "RAGQueryOptions with an explicit top_k of 0 serializes");
+    g_dummy_llm_last_top_k = -1;
+    rac_proto_buffer_init(&out);
+    rc = rac_rag_query_proto(session, explicit_zero_bytes.data(), explicit_zero_bytes.size(), &out);
+    CHECK(rc == RAC_SUCCESS, "RAG query with an explicit top_k of 0 succeeds");
+    CHECK(g_dummy_llm_last_top_k == 0, "RAG query honours an explicit top_k of 0");
+    rac_proto_buffer_free(&out);
+    rac_sdk_event_clear_queue();
+
     // A token-limited thinking phase may omit its closing tag. The RAG result
     // must keep that private content out of answer while retaining typed
     // thinking_content for non-UI consumers.