refactor: Upgrade agent-framework and azure-ai-projects libraries with API changes

infra/scripts/agent_scripts/01_create_agents.py

@@ -6,7 +6,7 @@
 from azure.identity.aio import AzureCliCredential
 from azure.ai.projects.models import (
     PromptAgentDefinition,
-    AzureAISearchAgentTool,
+    AzureAISearchTool,
     FunctionTool,
     AzureAISearchToolResource,
     AISearchIndexResource,
@@ -115,7 +115,7 @@ async def main():
                         }
                     ),
                     # Azure AI Search - built-in service tool (no client implementation needed)
-                    AzureAISearchAgentTool(
+                    AzureAISearchTool(
                         azure_ai_search=AzureAISearchToolResource(
                             indexes=[
                                 AISearchIndexResource(

infra/scripts/agent_scripts/requirements.txt

@@ -1,3 +1,3 @@
 aiohttp==3.13.4
 azure-identity==1.25.2
-azure-ai-projects==2.0.0b3
+azure-ai-projects==2.1.0

infra/scripts/index_scripts/03_cu_process_data_text.py

@@ -16,7 +16,7 @@
 
 # Suppress informational warnings from agent_framework about runtime
 # tool/structured_output overrides not being supported by AzureAIClient.
-logging.getLogger("agent_framework.azure").setLevel(logging.ERROR)
+logging.getLogger("agent_framework.foundry").setLevel(logging.ERROR)
 
 import pandas as pd
 import pyodbc
@@ -29,7 +29,7 @@
 from azure.search.documents.indexes import SearchIndexClient
 from azure.storage.filedatalake import DataLakeServiceClient
 
-from agent_framework.azure import AzureAIProjectAgentProvider
+from agent_framework_foundry import FoundryAgent
 
 from content_understanding_client import AzureContentUnderstandingClient, sanitize_cu_output
 
@@ -514,11 +514,8 @@ async def call_topic_mining_agent(topics_str1):
             AsyncAzureCliCredential(process_timeout=30) as async_cred,
             AIProjectClient(endpoint=AI_PROJECT_ENDPOINT, credential=async_cred) as project_client,
         ):
-            # Create provider for agent management
-            provider = AzureAIProjectAgentProvider(project_client=project_client)
-
-            # Get agent using provider
-            agent = await provider.get_agent(name=TOPIC_MINING_AGENT_NAME)
+            # Create agent using FoundryAgent
+            agent = FoundryAgent(project_client=project_client, agent_name=TOPIC_MINING_AGENT_NAME)
 
             # Query with the topics string
             query = f"Analyze these conversation topics and identify distinct categories: {topics_str1}"
@@ -561,11 +558,8 @@ async def map_all_topics():
             AsyncAzureCliCredential(process_timeout=30) as async_cred,
             AIProjectClient(endpoint=AI_PROJECT_ENDPOINT, credential=async_cred) as project_client,
         ):
-            # Create provider for agent management
-            provider = AzureAIProjectAgentProvider(project_client=project_client)
-
-            # Get agent using provider
-            agent = await provider.get_agent(name=TOPIC_MAPPING_AGENT_NAME)
+            # Create agent using FoundryAgent
+            agent = FoundryAgent(project_client=project_client, agent_name=TOPIC_MAPPING_AGENT_NAME)
 
             # Process all rows using the same agent instance
             for _, row in df_processed_data.iterrows():

infra/scripts/index_scripts/04_cu_process_custom_data.py

@@ -16,7 +16,7 @@
 
 # Suppress informational warnings from agent_framework about runtime
 # tool/structured_output overrides not being supported by AzureAIClient.
-logging.getLogger("agent_framework.azure").setLevel(logging.ERROR)
+logging.getLogger("agent_framework.foundry").setLevel(logging.ERROR)
 
 import pandas as pd
 import pyodbc
@@ -43,7 +43,7 @@
 )
 from azure.storage.filedatalake import DataLakeServiceClient
 
-from agent_framework.azure import AzureAIProjectAgentProvider
+from agent_framework_foundry import FoundryAgent
 
 from content_understanding_client import AzureContentUnderstandingClient, sanitize_cu_output
 
@@ -610,11 +610,8 @@ async def call_topic_mining_agent(topics_str1):
             AsyncAzureCliCredential(process_timeout=30) as async_cred,
             AIProjectClient(endpoint=AI_PROJECT_ENDPOINT, credential=async_cred) as project_client,
         ):
-            # Create provider for agent management
-            provider = AzureAIProjectAgentProvider(project_client=project_client)
-
-            # Get agent using provider
-            agent = await provider.get_agent(name=TOPIC_MINING_AGENT_NAME)
+            # Create agent using FoundryAgent
+            agent = FoundryAgent(project_client=project_client, agent_name=TOPIC_MINING_AGENT_NAME)
 
             # Query with the topics string
             query = f"Analyze these conversation topics and identify distinct categories: {topics_str1}"
@@ -657,11 +654,8 @@ async def map_all_topics():
             AsyncAzureCliCredential(process_timeout=30) as async_cred,
             AIProjectClient(endpoint=AI_PROJECT_ENDPOINT, credential=async_cred) as project_client,
         ):
-            # Create provider for agent management
-            provider = AzureAIProjectAgentProvider(project_client=project_client)
-
-            # Get agent using provider
-            agent = await provider.get_agent(name=TOPIC_MAPPING_AGENT_NAME)
+            # Create agent using FoundryAgent
+            agent = FoundryAgent(project_client=project_client, agent_name=TOPIC_MAPPING_AGENT_NAME)
 
             # Process all rows using the same agent instance
             for _, row in df_processed_data.iterrows():

infra/scripts/index_scripts/requirements.txt

@@ -1,9 +1,9 @@
 azure-storage-file-datalake==12.23.0
 openai==2.24.0
-azure-ai-projects==2.0.0b3
-azure-ai-agents==1.2.0b5
-agent-framework-core==1.0.0rc2
-agent-framework-azure-ai==1.0.0rc2
+azure-ai-projects==2.1.0
+azure-ai-inference==1.0.0b9
+agent-framework-core==1.3.0
+agent-framework-foundry==1.3.0
 pypdf==6.10.2
 tiktoken==0.12.0
 azure-identity==1.25.2

src/api/requirements.txt

@@ -14,10 +14,11 @@ aiohttp==3.13.5
 # Azure Services
 azure-identity==1.25.3
 azure-search-documents==11.6.0
-azure-ai-projects==2.0.0b3
-azure-ai-agents==1.2.0b5
-agent-framework-core==1.0.0rc2
-agent-framework-azure-ai==1.0.0rc2
+azure-ai-projects==2.1.0
+azure-ai-inference==1.0.0b9
+agent-framework-core==1.3.0
+agent-framework-foundry==1.3.0
+agent-framework-openai==1.3.0
 azure-cosmos==4.15.0
 
 # Additional utilities

src/api/services/_patches/__init__.py

@@ -0,0 +1 @@
+"""Runtime patches applied to third-party packages used by this app."""

src/api/services/_patches/agent_framework_search_citations.py

@@ -0,0 +1,175 @@
+"""Restore Azure AI Search ``get_url`` enrichment on streaming citations.
+
+Pre-GA ``agent-framework-azure-ai==1.0.0rc2`` enriched ``url_citation``
+streaming annotations with a per-document REST URL exposed under
+``additional_properties.get_url``. That subclass was removed when the
+``azure-ai`` package was retired at GA.
+
+In the GA ``agent_framework_openai._chat_client._parse_chunk_from_openai``:
+
+1. ``response.azure_ai_search_call_output.done`` events fall through to
+   the default ``case _:`` debug log, so the ``output.get_urls[]`` array
+   carrying per-document REST URLs is silently dropped.
+2. ``url_citation`` annotations (added by upstream PR #5071) emit only
+   ``title`` and ``url`` (the search-service root URL).
+
+This patch wraps that method to:
+
+1. Cache ``get_urls`` per-stream when seeing the search-call-output event.
+2. Inject ``additional_properties.get_url`` on ``url_citation`` annotations
+   using ``annotation_index`` as the lookup key, matching the pre-GA
+   contract that the citation extraction in ``chat_service.py`` reads.
+
+Tracking upstream: https://github.com/microsoft/agent-framework/issues/5995
+Safe to remove once upstream ports the ``get_url`` enrichment into
+``agent_framework_openai`` or ``agent_framework_foundry``.
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+from typing import Any
+
+_DOC_INDEX_RE = re.compile(r"^doc_(\d+)$")
+
+logger = logging.getLogger(__name__)
+
+_PATCH_APPLIED = False
+_CACHE_ATTR = "_kmsa_search_get_urls_cache"
+_TARGET_METHOD = "_parse_chunk_from_openai"
+_TARGET_CLASS = "RawOpenAIChatClient"
+_UPSTREAM_ISSUE = "https://github.com/microsoft/agent-framework/issues/5995"
+
+
+def apply() -> None:
+    """Idempotently patch RawOpenAIChatClient._parse_chunk_from_openai.
+
+    Safe to call multiple times; the second call is a no-op.
+    Logs a warning (does not raise) if upstream has renamed the target,
+    so app startup still succeeds with degraded citations.
+    """
+    global _PATCH_APPLIED
+    if _PATCH_APPLIED:
+        return
+
+    try:
+        from agent_framework_openai import _chat_client as _cc
+    except ImportError:
+        logger.warning(
+            "agent_framework_openai not installed; "
+            "Azure AI Search citation patch skipped"
+        )
+        return
+
+    target_cls = getattr(_cc, _TARGET_CLASS, None)
+    if target_cls is None or not hasattr(target_cls, _TARGET_METHOD):
+        logger.warning(
+            "agent-framework upgrade broke citation patch: %s.%s no longer exists. "
+            "Per-document URLs (get_url) will be missing on Azure AI Search "
+            "citations. See %s",
+            _TARGET_CLASS, _TARGET_METHOD, _UPSTREAM_ISSUE,
+        )
+        return
+
+    _original = getattr(target_cls, _TARGET_METHOD)
+
+    def _patched(self: Any, event: Any, *args: Any, **kwargs: Any) -> Any:
+        event_type = getattr(event, "type", None)
+
+        # Reset per-stream cache so back-to-back requests on the same client
+        # instance don't cross-pollute citation enrichment.
+        if event_type in ("response.created", "response.in_progress"):
+            setattr(self, _CACHE_ATTR, [])
+
+        # Capture get_urls from azure_ai_search_call_output items on .done.
+        # The .added event for this item has output='[]'; the .done event
+        # carries the actual documents + get_urls.
+        if event_type == "response.output_item.done":
+            try:
+                done_item = getattr(event, "item", None)
+                if getattr(done_item, "type", None) == "azure_ai_search_call_output":
+                    output = getattr(done_item, "output", None)
+                    get_urls = getattr(output, "get_urls", None)
+                    if get_urls is None and isinstance(output, dict):
+                        get_urls = output.get("get_urls")
+                    if get_urls is None and isinstance(output, str):
+                        # Some SDK versions deliver `output` as a JSON string.
+                        import json as _json
+                        try:
+                            parsed = _json.loads(output)
+                            if isinstance(parsed, dict):
+                                get_urls = parsed.get("get_urls")
+                        except Exception:  # noqa: BLE001
+                            pass
+                    if get_urls:
+                        cache = getattr(self, _CACHE_ATTR, None)
+                        if cache is None:
+                            cache = []
+                            setattr(self, _CACHE_ATTR, cache)
+                        cache.extend(get_urls)
+            except Exception:  # noqa: BLE001 - defensive: never break streaming
+                logger.debug(
+                    "search-citation patch: failed to capture get_urls",
+                    exc_info=True,
+                )
+
+        result = _original(self, event, *args, **kwargs)
+
+        # Enrich url_citation annotations emitted by the base method's
+        # response.output_text.annotation.added branch.
+        if event_type == "response.output_text.annotation.added":
+            try:
+                cache = getattr(self, _CACHE_ATTR, None) or []
+                if cache:
+                    for content in (getattr(result, "contents", None) or []):
+                        for ann in (getattr(content, "annotations", None) or []):
+                            if not isinstance(ann, dict):
+                                continue
+                            if ann.get("type") != "citation":
+                                continue
+                            add_props = ann.get("additional_properties") or {}
+                            # Idempotent: do not overwrite if upstream ever ships
+                            # the fix and starts populating get_url itself.
+                            if add_props.get("get_url"):
+                                continue
+                            # Map by title "doc_<N>" where N is the index into
+                            # the search results (and thus into get_urls). The
+                            # model can cite the same doc multiple times, so
+                            # annotation_index (a running counter) is unreliable.
+                            title = ann.get("title") or ""
+                            m = _DOC_INDEX_RE.match(str(title))
+                            doc_idx = int(m.group(1)) if m else None
+                            if doc_idx is None:
+                                doc_idx = add_props.get("annotation_index")
+                            if isinstance(doc_idx, int) and 0 <= doc_idx < len(cache):
+                                add_props["get_url"] = cache[doc_idx]
+                                ann["additional_properties"] = add_props
+            except Exception:  # noqa: BLE001
+                logger.debug(
+                    "search-citation patch: failed to enrich annotation",
+                    exc_info=True,
+                )
+
+        # Release per-stream state once the response completes.
+        if event_type == "response.completed":
+            try:
+                if hasattr(self, _CACHE_ATTR):
+                    delattr(self, _CACHE_ATTR)
+            except Exception:  # noqa: BLE001
+                pass
+
+        return result
+
+    setattr(target_cls, _TARGET_METHOD, _patched)
+    _PATCH_APPLIED = True
+    logger.info(
+        "Applied Azure AI Search citation patch on %s.%s (workaround for %s)",
+        _TARGET_CLASS, _TARGET_METHOD, _UPSTREAM_ISSUE,
+    )
+
+
+# Apply on import so a single
+# `import services._patches.agent_framework_search_citations`
+# from chat_service.py is enough.
+apply()