Upgrade to latest version of azure-search-documents and agentic retrieval API

app/backend/app.py

@@ -471,7 +471,6 @@ async def setup_clients():
     USE_CHAT_HISTORY_BROWSER = os.getenv("USE_CHAT_HISTORY_BROWSER", "").lower() == "true"
     USE_CHAT_HISTORY_COSMOS = os.getenv("USE_CHAT_HISTORY_COSMOS", "").lower() == "true"
     USE_AGENTIC_RETRIEVAL = os.getenv("USE_AGENTIC_RETRIEVAL", "").lower() == "true"
-    ENABLE_AGENTIC_RETRIEVAL_SOURCE_DATA = os.getenv("ENABLE_AGENTIC_RETRIEVAL_SOURCE_DATA", "").lower() == "true"
 
     # WEBSITE_HOSTNAME is always set by App Service, RUNNING_IN_PRODUCTION is set in main.bicep
     RUNNING_ON_AZURE = os.getenv("WEBSITE_HOSTNAME") is not None or os.getenv("RUNNING_IN_PRODUCTION") is not None
@@ -690,7 +689,6 @@ async def setup_clients():
         query_speller=AZURE_SEARCH_QUERY_SPELLER,
         prompt_manager=prompt_manager,
         reasoning_effort=OPENAI_REASONING_EFFORT,
-        hydrate_references=ENABLE_AGENTIC_RETRIEVAL_SOURCE_DATA,
         multimodal_enabled=USE_MULTIMODAL,
         image_embeddings_client=image_embeddings_client,
         global_blob_manager=global_blob_manager,
@@ -718,7 +716,6 @@ async def setup_clients():
         query_speller=AZURE_SEARCH_QUERY_SPELLER,
         prompt_manager=prompt_manager,
         reasoning_effort=OPENAI_REASONING_EFFORT,
-        hydrate_references=ENABLE_AGENTIC_RETRIEVAL_SOURCE_DATA,
         multimodal_enabled=USE_MULTIMODAL,
         image_embeddings_client=image_embeddings_client,
         global_blob_manager=global_blob_manager,

app/backend/approaches/approach.py

@@ -6,13 +6,13 @@
 
 from azure.search.documents.agent.aio import KnowledgeAgentRetrievalClient
 from azure.search.documents.agent.models import (
-    KnowledgeAgentAzureSearchDocReference,
-    KnowledgeAgentIndexParams,
     KnowledgeAgentMessage,
     KnowledgeAgentMessageTextContent,
     KnowledgeAgentRetrievalRequest,
     KnowledgeAgentRetrievalResponse,
-    KnowledgeAgentSearchActivityRecord,
+    KnowledgeAgentSearchIndexActivityRecord,
+    KnowledgeAgentSearchIndexReference,
+    SearchIndexKnowledgeSourceParams,
 )
 from azure.search.documents.aio import SearchClient
 from azure.search.documents.models import (
@@ -162,7 +162,6 @@ def __init__(
         openai_host: str,
         prompt_manager: PromptManager,
         reasoning_effort: Optional[str] = None,
-        hydrate_references: bool = False,
         multimodal_enabled: bool = False,
         image_embeddings_client: Optional[ImageEmbeddings] = None,
         global_blob_manager: Optional[BlobManager] = None,
@@ -180,7 +179,6 @@ def __init__(
         self.openai_host = openai_host
         self.prompt_manager = prompt_manager
         self.reasoning_effort = reasoning_effort
-        self.hydrate_references = hydrate_references
         self.include_token_usage = True
         self.multimodal_enabled = multimodal_enabled
         self.image_embeddings_client = image_embeddings_client
@@ -276,7 +274,6 @@ async def run_agentic_retrieval(
         top: Optional[int] = None,
         filter_add_on: Optional[str] = None,
         minimum_reranker_score: Optional[float] = None,
-        max_docs_for_reranker: Optional[int] = None,
         results_merge_strategy: Optional[str] = None,
     ) -> tuple[KnowledgeAgentRetrievalResponse, list[Document]]:
         # STEP 1: Invoke agentic retrieval
@@ -289,13 +286,10 @@ async def run_agentic_retrieval(
                     for msg in messages
                     if msg["role"] != "system"
                 ],
-                target_index_params=[
-                    KnowledgeAgentIndexParams(
-                        index_name=search_index_name,
-                        reranker_threshold=minimum_reranker_score,
-                        max_docs_for_reranker=max_docs_for_reranker,
+                knowledge_source_params=[
+                    SearchIndexKnowledgeSourceParams(
+                        knowledge_source_name=search_index_name,
                         filter_add_on=filter_add_on,
-                        include_reference_source_data=True,
                     )
                 ],
             )
@@ -305,12 +299,12 @@ async def run_agentic_retrieval(
         activities = response.activity
         activity_mapping: dict[int, str] = (
             {
-                activity.id: activity.query.search
+                activity.id: activity.search_index_arguments.search
                 for activity in activities
                 if (
-                    isinstance(activity, KnowledgeAgentSearchActivityRecord)
-                    and activity.query
-                    and activity.query.search is not None
+                    isinstance(activity, KnowledgeAgentSearchIndexActivityRecord)
+                    and activity.search_index_arguments
+                    and activity.search_index_arguments.search is not None
                 )
             }
             if activities
@@ -322,92 +316,42 @@ async def run_agentic_retrieval(
             return response, []
 
         # Extract references
-        refs = [r for r in response.references if isinstance(r, KnowledgeAgentAzureSearchDocReference)]
-
+        refs = [r for r in response.references if isinstance(r, KnowledgeAgentSearchIndexReference)]
         documents: list[Document] = []
-
-        if self.hydrate_references:
-            # Hydrate references to get full documents
-            documents = await self.hydrate_agent_references(
-                references=refs,
-                top=top,
-            )
-        else:
-            # Create documents from reference source data
-            for ref in refs:
-                if ref.source_data:
-                    documents.append(
-                        Document(
-                            id=ref.doc_key,
-                            content=ref.source_data.get("content"),
-                            sourcepage=ref.source_data.get("sourcepage"),
-                        )
-                    )
-                    if top and len(documents) >= top:
-                        break
-
-        # Build mappings for agent queries and sorting
-        ref_to_activity: dict[str, int] = {}
         doc_to_ref_id: dict[str, str] = {}
+
+        # Create documents from reference source data
         for ref in refs:
-            if ref.doc_key:
-                ref_to_activity[ref.doc_key] = ref.activity_source
+            if ref.source_data and ref.doc_key:
+                # Note that ref.doc_key is the same as source_data["id"]
+                documents.append(
+                    Document(
+                        id=ref.source_data.get("id"),
+                        content=ref.source_data.get("content"),
+                        category=ref.source_data.get("category"),
+                        sourcepage=ref.source_data.get("sourcepage"),
+                        sourcefile=ref.source_data.get("sourcefile"),
+                        oids=ref.source_data.get("oids"),
+                        groups=ref.source_data.get("groups"),
+                        reranker_score=ref.reranker_score,
+                        images=ref.source_data.get("images"),
+                        search_agent_query=activity_mapping[ref.activity_source],
+                    )
+                )
                 doc_to_ref_id[ref.doc_key] = ref.id
+                if top and len(documents) >= top:
+                    break
 
-        # Inject agent search queries into all documents
-        for doc in documents:
-            if doc.id and doc.id in ref_to_activity:
-                activity_id = ref_to_activity[doc.id]
-                doc.search_agent_query = activity_mapping.get(activity_id, "")
+        if minimum_reranker_score is not None:
+            documents = [doc for doc in documents if (doc.reranker_score or 0) >= minimum_reranker_score]
 
-        # Apply sorting strategy to the documents
-        if results_merge_strategy == "interleaved":  # Use interleaved reference order
+        if results_merge_strategy == "interleaved":
             documents = sorted(
                 documents,
                 key=lambda d: int(doc_to_ref_id.get(d.id, 0)) if d.id and doc_to_ref_id.get(d.id) else 0,
             )
-        # else: Default - preserve original order
-
         return response, documents
 
-    async def hydrate_agent_references(
-        self,
-        references: list[KnowledgeAgentAzureSearchDocReference],
-        top: Optional[int],
-    ) -> list[Document]:
-        doc_keys: set[str] = set()
-
-        for ref in references:
-            if not ref.doc_key:
-                continue
-            doc_keys.add(ref.doc_key)
-            if top and len(doc_keys) >= top:
-                break
-
-        if not doc_keys:
-            return []
-
-        # Build search filter only on unique doc IDs
-        id_csv = ",".join(doc_keys)
-        id_filter = f"search.in(id, '{id_csv}', ',')"
-
-        # Fetch full documents
-        hydrated_docs: list[Document] = await self.search(
-            top=len(doc_keys),
-            query_text=None,
-            filter=id_filter,
-            vectors=[],
-            use_text_search=False,
-            use_vector_search=False,
-            use_semantic_ranker=False,
-            use_semantic_captions=False,
-            minimum_search_score=None,
-            minimum_reranker_score=None,
-            use_query_rewriting=False,
-        )
-
-        return hydrated_docs
-
     async def get_sources_content(
         self,
         results: list[Document],

app/backend/approaches/chatreadretrieveread.py

@@ -56,7 +56,6 @@ def __init__(
         query_speller: str,
         prompt_manager: PromptManager,
         reasoning_effort: Optional[str] = None,
-        hydrate_references: bool = False,
         multimodal_enabled: bool = False,
         image_embeddings_client: Optional[ImageEmbeddings] = None,
         global_blob_manager: Optional[BlobManager] = None,
@@ -84,7 +83,6 @@ def __init__(
         self.query_rewrite_tools = self.prompt_manager.load_tools("chat_query_rewrite_tools.json")
         self.answer_prompt = self.prompt_manager.load_prompt("chat_answer_question.prompty")
         self.reasoning_effort = reasoning_effort
-        self.hydrate_references = hydrate_references
         self.include_token_usage = True
         self.multimodal_enabled = multimodal_enabled
         self.image_embeddings_client = image_embeddings_client
@@ -390,13 +388,10 @@ async def run_agentic_retrieval_approach(
         overrides: dict[str, Any],
         auth_claims: dict[str, Any],
     ):
-        minimum_reranker_score = overrides.get("minimum_reranker_score", 0)
         search_index_filter = self.build_filter(overrides, auth_claims)
+        minimum_reranker_score = overrides.get("minimum_reranker_score", 0)
         top = overrides.get("top", 3)
-        max_subqueries = overrides.get("max_subqueries", 10)
         results_merge_strategy = overrides.get("results_merge_strategy", "interleaved")
-        # 50 is the amount of documents that the reranker can process per query
-        max_docs_for_reranker = max_subqueries * 50
         send_text_sources = overrides.get("send_text_sources", True)
         send_image_sources = overrides.get("send_image_sources", self.multimodal_enabled) and self.multimodal_enabled
 
@@ -407,7 +402,6 @@ async def run_agentic_retrieval_approach(
             top=top,
             filter_add_on=search_index_filter,
             minimum_reranker_score=minimum_reranker_score,
-            max_docs_for_reranker=max_docs_for_reranker,
             results_merge_strategy=results_merge_strategy,
         )
 
@@ -426,7 +420,6 @@ async def run_agentic_retrieval_approach(
                     messages,
                     {
                         "reranker_threshold": minimum_reranker_score,
-                        "max_docs_for_reranker": max_docs_for_reranker,
                         "results_merge_strategy": results_merge_strategy,
                         "filter": search_index_filter,
                     },

app/backend/approaches/retrievethenread.py

@@ -46,7 +46,6 @@ def __init__(
         query_speller: str,
         prompt_manager: PromptManager,
         reasoning_effort: Optional[str] = None,
-        hydrate_references: bool = False,
         multimodal_enabled: bool = False,
         image_embeddings_client: Optional[ImageEmbeddings] = None,
         global_blob_manager: Optional[BlobManager] = None,
@@ -74,7 +73,6 @@ def __init__(
         self.answer_prompt = self.prompt_manager.load_prompt("ask_answer_question.prompty")
         self.reasoning_effort = reasoning_effort
         self.include_token_usage = True
-        self.hydrate_references = hydrate_references
         self.multimodal_enabled = multimodal_enabled
         self.image_embeddings_client = image_embeddings_client
         self.global_blob_manager = global_blob_manager
@@ -229,10 +227,7 @@ async def run_agentic_retrieval_approach(
         minimum_reranker_score = overrides.get("minimum_reranker_score", 0)
         search_index_filter = self.build_filter(overrides, auth_claims)
         top = overrides.get("top", 3)
-        max_subqueries = overrides.get("max_subqueries", 10)
         results_merge_strategy = overrides.get("results_merge_strategy", "interleaved")
-        # 50 is the amount of documents that the reranker can process per query
-        max_docs_for_reranker = max_subqueries * 50
         send_text_sources = overrides.get("send_text_sources", True)
         send_image_sources = overrides.get("send_image_sources", self.multimodal_enabled) and self.multimodal_enabled
 
@@ -243,7 +238,6 @@ async def run_agentic_retrieval_approach(
             top=top,
             filter_add_on=search_index_filter,
             minimum_reranker_score=minimum_reranker_score,
-            max_docs_for_reranker=max_docs_for_reranker,
             results_merge_strategy=results_merge_strategy,
         )
 
@@ -263,7 +257,6 @@ async def run_agentic_retrieval_approach(
                     messages,
                     {
                         "reranker_threshold": minimum_reranker_score,
-                        "max_docs_for_reranker": max_docs_for_reranker,
                         "results_merge_strategy": results_merge_strategy,
                         "filter": search_index_filter,
                     },

app/backend/prepdocslib/searchmanager.py

@@ -14,12 +14,14 @@
     KnowledgeAgent,
     KnowledgeAgentAzureOpenAIModel,
     KnowledgeAgentRequestLimits,
-    KnowledgeAgentTargetIndex,
+    KnowledgeSourceReference,
     RescoringOptions,
     SearchableField,
     SearchField,
     SearchFieldDataType,
     SearchIndex,
+    SearchIndexKnowledgeSource,
+    SearchIndexKnowledgeSourceParameters,
     SemanticConfiguration,
     SemanticField,
     SemanticPrioritizedFields,
@@ -83,7 +85,6 @@ async def create_index(self):
         logger.info("Checking whether search index %s exists...", self.search_info.index_name)
 
         async with self.search_info.create_search_index_client() as search_index_client:
-
             embedding_field = None
             images_field = None
             text_vector_search_profile = None
@@ -439,13 +440,29 @@ async def create_agent(self):
         if self.search_info.agent_name:
             logger.info(f"Creating search agent named {self.search_info.agent_name}")
 
+            field_names = ["id", "sourcepage", "sourcefile", "content", "category"]
+            if self.use_acls:
+                field_names.extend(["oids", "groups"])
+            if self.search_images:
+                field_names.append("images/url")
             async with self.search_info.create_search_index_client() as search_index_client:
+                knowledge_source = SearchIndexKnowledgeSource(
+                    name=self.search_info.index_name,  # Use the same name for convenience
+                    description="Default knowledge source using the main search index",
+                    search_index_parameters=SearchIndexKnowledgeSourceParameters(
+                        search_index_name=self.search_info.index_name,
+                        source_data_select=",".join(field_names),
+                    ),
+                )
+                await search_index_client.create_or_update_knowledge_source(
+                    knowledge_source=knowledge_source, api_version="2025-08-01-preview"
+                )
                 await search_index_client.create_or_update_agent(
                     agent=KnowledgeAgent(
                         name=self.search_info.agent_name,
-                        target_indexes=[
-                            KnowledgeAgentTargetIndex(
-                                index_name=self.search_info.index_name, default_include_reference_source_data=True
+                        knowledge_sources=[
+                            KnowledgeSourceReference(
+                                name=knowledge_source.name, include_references=True, include_reference_source_data=True
                             )
                         ],
                         models=[

app/backend/requirements.in

@@ -7,7 +7,7 @@ tenacity
 azure-ai-documentintelligence==1.0.0b4
 azure-cognitiveservices-speech
 azure-cosmos
-azure-search-documents==11.6.0b12
+azure-search-documents==11.7.0b1
 azure-storage-blob
 azure-storage-file-datalake
 uvicorn

app/backend/requirements.txt

@@ -56,7 +56,7 @@ azure-monitor-opentelemetry==1.6.13
     # via -r requirements.in
 azure-monitor-opentelemetry-exporter==1.0.0b40
     # via azure-monitor-opentelemetry
-azure-search-documents==11.6.0b12
+azure-search-documents==11.7.0b1
     # via -r requirements.in
 azure-storage-blob==12.22.0
     # via
@@ -122,7 +122,7 @@ hpack==4.1.0
     # via h2
 httpcore==1.0.9
     # via httpx
-httpx[http2]==0.27.0
+httpx==0.27.0
     # via
     #   microsoft-kiota-http
     #   msgraph-core
@@ -343,7 +343,7 @@ pydantic-core==2.20.1
     # via pydantic
 pygments==2.18.0
     # via rich
-pyjwt[crypto]==2.10.1
+pyjwt==2.10.1
     # via
     #   -r requirements.in
     #   msal