Add collaborativeagents module and update gitignore

- Add collaborativeagents subproject with adapters, agents, and evaluation modules
- Update .gitignore to exclude large binary files (.whl, .tar), wandb logs, and results

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

1 files changed, 26 insertions, 16 deletions

diff --git a/src/personalization/retrieval/pipeline.py b/src/personalization/retrieval/pipeline.py
index 3d3eeb7..e83940d 100644
--- a/src/personalization/retrieval/pipeline.py
+++ b/src/personalization/retrieval/pipeline.py
@@ -110,10 +110,14 @@ def retrieve_with_policy(
 
     candidates = [memory_cards[i] for i in dense_idx]
     candidate_docs = [c.note_text for c in candidates]
-    
+
     # 2. Rerank base score (P(yes|q,m))
-    base_scores = np.array(reranker.score(query, candidate_docs))
-    
+    # Skip reranking if we have fewer candidates than topk_rerank (saves GPU memory)
+    if len(candidates) <= topk_rerank:
+        base_scores = np.ones(len(candidates))  # Uniform scores
+    else:
+        base_scores = np.array(reranker.score(query, candidate_docs))
+
     # 3. Policy Scoring (Softmax)
     user_state: UserState = user_store.get_state(user_id)
     candidate_vectors = item_vectors[dense_idx] # [K, k]
@@ -181,29 +185,35 @@ def retrieve_no_policy(
 
     # 1. Dense retrieval
     dense_idx = dense_topk_indices(
-        query, 
-        embed_model, 
-        memory_embeddings, 
+        query,
+        embed_model,
+        memory_embeddings,
         valid_indices=valid_indices,
         topk=topk_dense
     )
-    
+
     if not dense_idx:
         return [], np.array([]), np.array([]), [], np.array([])
 
     candidates = [memory_cards[i] for i in dense_idx]
     candidate_docs = [c.note_text for c in candidates]
-    
+
     # 2. Rerank base score (P(yes|q,m))
-    base_scores = np.array(reranker.score(query, candidate_docs))
-    
-    # 3. Deterministic Top-K selection based on rerank scores ONLY (no policy)
-    k = min(topk_rerank, len(base_scores))
-    top_indices_local = base_scores.argsort()[-k:][::-1]
-    chosen_indices = top_indices_local.tolist()
-    
+    # Skip reranking if we have fewer candidates than topk_rerank (saves GPU memory)
+    if len(candidates) <= topk_rerank:
+        # Just return all candidates without reranking
+        base_scores = np.ones(len(candidates))  # Uniform scores
+        chosen_indices = list(range(len(candidates)))
+    else:
+        base_scores = np.array(reranker.score(query, candidate_docs))
+
+        # 3. Deterministic Top-K selection based on rerank scores ONLY (no policy)
+        k = min(topk_rerank, len(base_scores))
+        top_indices_local = base_scores.argsort()[-k:][::-1]
+        chosen_indices = top_indices_local.tolist()
+
     # Get scores for chosen items (for logging compatibility)
-    chosen_scores = base_scores[top_indices_local]
+    chosen_scores = base_scores[chosen_indices]
     
     # Return empty item vectors (not used in NoPersonal mode)
     # Return rerank scores as the "probs" field for logging compatibility