Remove Redundant Dedupe Logic (#1577)

backend/danswer/one_shot_answer/answer_question.py

@@ -208,15 +208,14 @@ def stream_answer_objects(
                     search_response_summary.top_sections
                 )
 
+                # Deduping happens at the last step to avoid harming quality by dropping content early on
                 deduped_docs = top_docs
                 if query_req.retrieval_options.dedupe_docs:
                     deduped_docs, dropped_inds = dedupe_documents(top_docs)
 
                 reference_db_search_docs = [
-                    create_db_search_doc(
-                        server_search_doc=top_doc, db_session=db_session
-                    )
-                    for top_doc in deduped_docs
+                    create_db_search_doc(server_search_doc=doc, db_session=db_session)
+                    for doc in deduped_docs
                 ]
 
                 response_docs = [

backend/danswer/search/models.py

@@ -78,9 +78,6 @@ class SearchRequest(ChunkContext):
     skip_rerank: bool | None = None
     skip_llm_chunk_filter: bool | None = None
 
-    # If this is set, only the highest matching chunk (or merged chunks) is returned
-    dedupe_docs: bool = False
-
     class Config:
         arbitrary_types_allowed = True
 
@@ -118,6 +115,8 @@ class RetrievalDetails(ChunkContext):
     # if None, no offset / limit
     offset: int | None = None
     limit: int | None = None
+
+    # If this is set, only the highest matching chunk (or merged chunks) is returned
     dedupe_docs: bool = False
 
 

backend/danswer/tools/search/search_tool.py

@@ -202,9 +202,6 @@ class SearchTool(Tool):
                 chunks_above=self.chunks_above,
                 chunks_below=self.chunks_below,
                 full_doc=self.full_doc,
-                dedupe_docs=self.retrieval_options.dedupe_docs
-                if self.retrieval_options
-                else False,
             ),
             user=self.user,
             db_session=self.db_session,