Merge pull request #1319 from JJordan0C/fix/search-graph-semantic-only-results

fix(mcp): isolate semantic-only JSON search
2026-11-09 09:11:00 +08:00 · 2026-08-18 00:04:36 +02:00
parent 0273250c33 27eab7362d
commit de4ec9b5e9
2 changed files with 76 additions and 38 deletions
@@ -3170,22 +3170,6 @@ static bool run_semantic_query_core(const char *args, cbm_store_t *store, const
    return type_error;
 }

-/* Append the semantic_query vector-search results onto the doc.  Returns
- * true if semantic_query was provided as a non-array (type error — caller
- * should surface to the user). */
-static bool run_semantic_query(yyjson_mut_doc *doc, yyjson_mut_val *root, const char *args,
-                               cbm_store_t *store, const char *project, int limit) {
-    cbm_vector_result_t *vresults = NULL;
-    int vcount = 0;
-    bool type_error =
-        run_semantic_query_core(args, store, project, limit, &vresults, &vcount, NULL);
-    if (vcount > 0) {
-        emit_semantic_results(doc, root, vresults, vcount);
-        cbm_store_free_vector_results(vresults, vcount);
-    }
-    return type_error;
-}
-
 /* ── Tree output for search_graph ───────────────────────────────────
 * Default response encoding: grouped tree rows (compact_out.h). The same
 * model is available as structured JSON via format:"json"; include_connected
@@ -3620,6 +3604,9 @@ static char *handle_search_graph(cbm_mcp_server_t *srv, const char *args) {
        .min_degree = min_degree,
        .max_degree = max_degree,
    };
+    bool has_structural_filters = label || name_pattern || qn_pattern || file_pattern ||
+                                  relationship || exclude_entry_points ||
+                                  min_degree != CBM_NOT_FOUND || max_degree != CBM_NOT_FOUND;

    if (!legacy_json) {
        const char *fields[SG_MAX_EXTRA_FIELDS];
@@ -3640,10 +3627,7 @@ static char *handle_search_graph(cbm_mcp_server_t *srv, const char *args) {
            /* Semantic-only calls get semantic results only: the legacy
             * behavior also ran the UNFILTERED regex search and prepended
             * up to `limit` unrelated enriched nodes to the response. */
-            bool has_filters = label || name_pattern || qn_pattern || file_pattern ||
-                               relationship || exclude_entry_points ||
-                               min_degree != CBM_NOT_FOUND || max_degree != CBM_NOT_FOUND;
-            bool semantic_only = sq_present && !has_filters;
+            bool semantic_only = sq_present && !has_structural_filters;

            cbm_sb_t sb;
            cbm_sb_init(&sb);
@@ -3727,8 +3711,31 @@ static char *handle_search_graph(cbm_mcp_server_t *srv, const char *args) {
            true);
    }

+    cbm_vector_result_t *vresults = NULL;
+    int vcount = 0;
+    bool sq_present = false;
+    bool sq_type_error =
+        run_semantic_query_core(args, store, project, limit, &vresults, &vcount, &sq_present);
+    if (sq_type_error) {
+        free(project);
+        free(label);
+        free(name_pattern);
+        free(qn_pattern);
+        free(file_pattern);
+        free(relationship);
+        return cbm_mcp_text_result(
+            "semantic_query must be an array of keyword strings, e.g. "
+            "[\"send\",\"pubsub\",\"publish\"] — not a single string. Split your query "
+            "into individual keywords; each is scored independently via per-keyword "
+            "min-cosine.",
+            true);
+    }
+
+    bool semantic_only = sq_present && !has_structural_filters;
    cbm_search_output_t out = {0};
-    cbm_store_search(store, &params, &out);
+    if (!semantic_only) {
+        cbm_store_search(store, &params, &out);
+    }

    yyjson_mut_doc *doc = yyjson_mut_doc_new(NULL);
    yyjson_mut_val *root = yyjson_mut_obj(doc);
@@ -3767,23 +3774,9 @@ static char *handle_search_graph(cbm_mcp_server_t *srv, const char *args) {
        }
    }

-    bool sq_type_error = run_semantic_query(doc, root, args, store, project, limit);
-
-    if (sq_type_error) {
-        yyjson_mut_doc_free(doc);
-        cbm_store_search_free(&out);
-        free(project);
-        free(label);
-        free(name_pattern);
-        free(qn_pattern);
-        free(file_pattern);
-        free(relationship);
-        return cbm_mcp_text_result(
-            "semantic_query must be an array of keyword strings, e.g. "
-            "[\"send\",\"pubsub\",\"publish\"] — not a single string. Split your query "
-            "into individual keywords; each is scored independently via per-keyword "
-            "min-cosine.",
-            true);
+    if (vcount > 0) {
+        emit_semantic_results(doc, root, vresults, vcount);
+        cbm_store_free_vector_results(vresults, vcount);
    }

    char *json = yy_doc_to_str(doc);
@@ -1637,6 +1637,50 @@ static cbm_mcp_server_t *setup_snippet_server(char *tmp_dir, size_t tmp_sz);
 static void cleanup_snippet_dir(const char *tmp_dir);
 static char *extract_text_content(const char *mcp_result);

+TEST(tool_search_graph_semantic_only_skips_structural_results_issue1295) {
+    cbm_mcp_server_t *srv = cbm_mcp_server_new(NULL);
+    ASSERT_NOT_NULL(srv);
+    cbm_store_t *st = cbm_mcp_server_store(srv);
+    const char *proj = "semantic-only";
+    cbm_mcp_server_set_project(srv, proj);
+    cbm_store_upsert_project(st, proj, "/tmp/semantic-only");
+
+    cbm_node_t unrelated = {.project = proj,
+                            .label = "Function",
+                            .name = "unrelated_node",
+                            .qualified_name = "semantic-only.unrelated_node",
+                            .file_path = "unrelated.c",
+                            .start_line = 1,
+                            .end_line = 2};
+    ASSERT_GT(cbm_store_upsert_node(st, &unrelated), 0);
+
+    char *resp = cbm_mcp_handle_tool(
+        srv, "search_graph",
+        "{\"project\":\"semantic-only\",\"semantic_query\":[\"publish\"],\"limit\":5}");
+    ASSERT_NOT_NULL(resp);
+    char *inner = extract_text_content(resp);
+    free(resp);
+    ASSERT_NOT_NULL(inner);
+    ASSERT_NULL(strstr(inner, "unrelated_node"));
+    free(inner);
+
+    resp = cbm_mcp_handle_tool(srv, "search_graph",
+                               "{\"project\":\"semantic-only\",\"semantic_query\":[\"publish\"],"
+                               "\"format\":\"json\",\"limit\":5}");
+    ASSERT_NOT_NULL(resp);
+    inner = extract_text_content(resp);
+    free(resp);
+    ASSERT_NOT_NULL(inner);
+    ASSERT_NOT_NULL(strstr(inner, "\"total\":0"));
+    ASSERT_NOT_NULL(strstr(inner, "\"count\":0"));
+    ASSERT_NOT_NULL(strstr(inner, "\"groups\":[]"));
+    ASSERT_NULL(strstr(inner, "unrelated_node"));
+    free(inner);
+
+    cbm_mcp_server_free(srv);
+    PASS();
+}
+
 /* callers_total/callees_total must count what the caller can enumerate: with
 * include_tests=false (default) test-file rows are hidden from the table, so
 * the totals must apply the same filter — a raw visited_count overstated the
@@ -10815,6 +10859,7 @@ SUITE(mcp) {
    RUN_TEST(tool_get_graph_schema_empty);
    RUN_TEST(tool_unknown_tool);
    RUN_TEST(tool_search_graph_basic);
+    RUN_TEST(tool_search_graph_semantic_only_skips_structural_results_issue1295);
    RUN_TEST(tool_trace_totals_respect_test_filter);
    RUN_TEST(tool_trace_totals_respect_test_filter_tests_root_subtree_issue1294);
    RUN_TEST(tool_get_architecture_cycles_detects_scc);