sort?

app.py

@@ -188,7 +188,7 @@ def load_corpus_lookups(dataset_name):
         corpus_lookups[dataset_name] += p_lookup
 
     logger.info(f"Loaded corpus lookups for {dataset_name}. Total entries: {len(corpus_lookups[dataset_name])}")
-    logger.info(f"Sample corpus lookup entry: {corpus_lookups[dataset_name][0]}")
+    logger.info(f"Sample corpus lookup entry: {corpus_lookups[dataset_name][:10]}")
 
 def load_queries(dataset_name):
     global queries, q_lookups, qrels
@@ -244,8 +244,23 @@ def run_evaluation(dataset, postfix):
 
     all_scores, psg_indices = search_queries(dataset, q_reps)
     
-    results = {str(qid): {str(doc_id): float(score) for doc_id, score in zip(doc_ids, scores)}
-           for qid, scores, doc_ids in zip(q_lookups[dataset].keys(), all_scores, psg_indices)}
+    results = {}
+    logging.info(f"Number of queries in q_lookups: {len(q_lookups[dataset])}")
+    logging.info("Size of all_scores: " + str(len(all_scores)))
+    logging.info("Size of psg_indices: " + str(len(psg_indices)))
+    for qid, scores, doc_ids in zip(q_lookups[dataset].keys(), all_scores, psg_indices):
+        qid_str = str(qid)
+        results[qid_str] = {}
+        for doc_id, score in zip(doc_ids, scores):
+            doc_id_str = str(doc_id)
+            results[qid_str][doc_id_str] = float(score)
+
+        if not results[qid_str]:  # If no results for this query
+            logger.warning(f"No results for query {qid_str}")
+
+    logger.info(f"Number of queries in results: {len(results)}")
+    logger.info(f"Sample result: {next(iter(results.items()))}")
+
     qrels[dataset] = {str(qid): {str(doc_id): rel for doc_id, rel in rels.items()}
                   for qid, rels in qrels[dataset].items()}