Merge branch 'main' into fix/issue-2584-llama4-tool-calling

tests/integration/agents/test_agents.py

@@ -77,6 +77,24 @@ def agent_config(llama_stack_client, text_model_id):
     return agent_config
 
 
+@pytest.fixture(scope="session")
+def agent_config_without_safety(text_model_id):
+    agent_config = dict(
+        model=text_model_id,
+        instructions="You are a helpful assistant",
+        sampling_params={
+            "strategy": {
+                "type": "top_p",
+                "temperature": 0.0001,
+                "top_p": 0.9,
+            },
+        },
+        tools=[],
+        enable_session_persistence=False,
+    )
+    return agent_config
+
+
 def test_agent_simple(llama_stack_client, agent_config):
     agent = Agent(llama_stack_client, **agent_config)
     session_id = agent.create_session(f"test-session-{uuid4()}")
@@ -491,7 +509,7 @@ def test_rag_agent(llama_stack_client, agent_config, rag_tool_name):
             assert expected_kw in response.output_message.content.lower()
 
 
-def test_rag_agent_with_attachments(llama_stack_client, agent_config):
+def test_rag_agent_with_attachments(llama_stack_client, agent_config_without_safety):
     urls = ["llama3.rst", "lora_finetune.rst"]
     documents = [
         # passign as url
@@ -514,14 +532,8 @@ def test_rag_agent_with_attachments(llama_stack_client, agent_config):
             metadata={},
         ),
     ]
-    rag_agent = Agent(llama_stack_client, **agent_config)
+    rag_agent = Agent(llama_stack_client, **agent_config_without_safety)
     session_id = rag_agent.create_session(f"test-session-{uuid4()}")
-    user_prompts = [
-        (
-            "Instead of the standard multi-head attention, what attention type does Llama3-8B use?",
-            "grouped",
-        ),
-    ]
     user_prompts = [
         (
             "I am attaching some documentation for Torchtune. Help me answer questions I will ask next.",
@@ -549,82 +561,6 @@ def test_rag_agent_with_attachments(llama_stack_client, agent_config):
     assert "lora" in response.output_message.content.lower()
 
 
-@pytest.mark.skip(reason="Code interpreter is currently disabled in the Stack")
-def test_rag_and_code_agent(llama_stack_client, agent_config):
-    if "llama-4" in agent_config["model"].lower():
-        pytest.xfail("Not working for llama4")
-
-    documents = []
-    documents.append(
-        Document(
-            document_id="nba_wiki",
-            content="The NBA was created on August 3, 1949, with the merger of the Basketball Association of America (BAA) and the National Basketball League (NBL).",
-            metadata={},
-        )
-    )
-    documents.append(
-        Document(
-            document_id="perplexity_wiki",
-            content="""Perplexity the company was founded in 2022 by Aravind Srinivas, Andy Konwinski, Denis Yarats and Johnny Ho, engineers with backgrounds in back-end systems, artificial intelligence (AI) and machine learning:
-
-    Srinivas, the CEO, worked at OpenAI as an AI researcher.
-    Konwinski was among the founding team at Databricks.
-    Yarats, the CTO, was an AI research scientist at Meta.
-    Ho, the CSO, worked as an engineer at Quora, then as a quantitative trader on Wall Street.[5]""",
-            metadata={},
-        )
-    )
-    vector_db_id = f"test-vector-db-{uuid4()}"
-    llama_stack_client.vector_dbs.register(
-        vector_db_id=vector_db_id,
-        embedding_model="all-MiniLM-L6-v2",
-        embedding_dimension=384,
-    )
-    llama_stack_client.tool_runtime.rag_tool.insert(
-        documents=documents,
-        vector_db_id=vector_db_id,
-        chunk_size_in_tokens=128,
-    )
-    agent_config = {
-        **agent_config,
-        "tools": [
-            dict(
-                name="builtin::rag/knowledge_search",
-                args={"vector_db_ids": [vector_db_id]},
-            ),
-            "builtin::code_interpreter",
-        ],
-    }
-    agent = Agent(llama_stack_client, **agent_config)
-    user_prompts = [
-        (
-            "when was Perplexity the company founded?",
-            [],
-            "knowledge_search",
-            "2022",
-        ),
-        (
-            "when was the nba created?",
-            [],
-            "knowledge_search",
-            "1949",
-        ),
-    ]
-
-    for prompt, docs, tool_name, expected_kw in user_prompts:
-        session_id = agent.create_session(f"test-session-{uuid4()}")
-        response = agent.create_turn(
-            messages=[{"role": "user", "content": prompt}],
-            session_id=session_id,
-            documents=docs,
-            stream=False,
-        )
-        tool_execution_step = next(step for step in response.steps if step.step_type == "tool_execution")
-        assert tool_execution_step.tool_calls[0].tool_name == tool_name, f"Failed on {prompt}"
-        if expected_kw:
-            assert expected_kw in response.output_message.content.lower()
-
-
 @pytest.mark.parametrize(
     "client_tools",
     [(get_boiling_point, False), (get_boiling_point_with_metadata, True)],

tests/unit/providers/vector_io/remote/test_milvus.py

@@ -0,0 +1,191 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from unittest.mock import MagicMock, patch
+
+import numpy as np
+import pytest
+import pytest_asyncio
+
+from llama_stack.apis.vector_io import QueryChunksResponse
+
+# Mock the entire pymilvus module
+pymilvus_mock = MagicMock()
+pymilvus_mock.DataType = MagicMock()
+pymilvus_mock.MilvusClient = MagicMock
+
+# Apply the mock before importing MilvusIndex
+with patch.dict("sys.modules", {"pymilvus": pymilvus_mock}):
+    from llama_stack.providers.remote.vector_io.milvus.milvus import MilvusIndex
+
+# This test is a unit test for the MilvusVectorIOAdapter class. This should only contain
+# tests which are specific to this class. More general (API-level) tests should be placed in
+# tests/integration/vector_io/
+#
+# How to run this test:
+#
+# pytest tests/unit/providers/vector_io/test_milvus.py \
+# -v -s --tb=short --disable-warnings --asyncio-mode=auto
+
+MILVUS_PROVIDER = "milvus"
+
+
+@pytest_asyncio.fixture
+async def mock_milvus_client() -> MagicMock:
+    """Create a mock Milvus client with common method behaviors."""
+    client = MagicMock()
+
+    # Mock collection operations
+    client.has_collection.return_value = False  # Initially no collection
+    client.create_collection.return_value = None
+    client.drop_collection.return_value = None
+
+    # Mock insert operation
+    client.insert.return_value = {"insert_count": 10}
+
+    # Mock search operation - return mock results (data should be dict, not JSON string)
+    client.search.return_value = [
+        [
+            {
+                "id": 0,
+                "distance": 0.1,
+                "entity": {"chunk_content": {"content": "mock chunk 1", "metadata": {"document_id": "doc1"}}},
+            },
+            {
+                "id": 1,
+                "distance": 0.2,
+                "entity": {"chunk_content": {"content": "mock chunk 2", "metadata": {"document_id": "doc2"}}},
+            },
+        ]
+    ]
+
+    # Mock query operation for keyword search (data should be dict, not JSON string)
+    client.query.return_value = [
+        {
+            "chunk_id": "chunk1",
+            "chunk_content": {"content": "mock chunk 1", "metadata": {"document_id": "doc1"}},
+            "score": 0.9,
+        },
+        {
+            "chunk_id": "chunk2",
+            "chunk_content": {"content": "mock chunk 2", "metadata": {"document_id": "doc2"}},
+            "score": 0.8,
+        },
+        {
+            "chunk_id": "chunk3",
+            "chunk_content": {"content": "mock chunk 3", "metadata": {"document_id": "doc3"}},
+            "score": 0.7,
+        },
+    ]
+
+    return client
+
+
+@pytest_asyncio.fixture
+async def milvus_index(mock_milvus_client):
+    """Create a MilvusIndex with mocked client."""
+    index = MilvusIndex(client=mock_milvus_client, collection_name="test_collection")
+    yield index
+    # No real cleanup needed since we're using mocks
+
+
+@pytest.mark.asyncio
+async def test_add_chunks(milvus_index, sample_chunks, sample_embeddings, mock_milvus_client):
+    # Setup: collection doesn't exist initially, then exists after creation
+    mock_milvus_client.has_collection.side_effect = [False, True]
+
+    await milvus_index.add_chunks(sample_chunks, sample_embeddings)
+
+    # Verify collection was created and data was inserted
+    mock_milvus_client.create_collection.assert_called_once()
+    mock_milvus_client.insert.assert_called_once()
+
+    # Verify the insert call had the right number of chunks
+    insert_call = mock_milvus_client.insert.call_args
+    assert len(insert_call[1]["data"]) == len(sample_chunks)
+
+
+@pytest.mark.asyncio
+async def test_query_chunks_vector(
+    milvus_index, sample_chunks, sample_embeddings, embedding_dimension, mock_milvus_client
+):
+    # Setup: Add chunks first
+    mock_milvus_client.has_collection.return_value = True
+    await milvus_index.add_chunks(sample_chunks, sample_embeddings)
+
+    # Test vector search
+    query_embedding = np.random.rand(embedding_dimension).astype(np.float32)
+    response = await milvus_index.query_vector(query_embedding, k=2, score_threshold=0.0)
+
+    assert isinstance(response, QueryChunksResponse)
+    assert len(response.chunks) == 2
+    mock_milvus_client.search.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_query_chunks_keyword_search(milvus_index, sample_chunks, sample_embeddings, mock_milvus_client):
+    mock_milvus_client.has_collection.return_value = True
+    await milvus_index.add_chunks(sample_chunks, sample_embeddings)
+
+    # Test keyword search
+    query_string = "Sentence 5"
+    response = await milvus_index.query_keyword(query_string=query_string, k=2, score_threshold=0.0)
+
+    assert isinstance(response, QueryChunksResponse)
+    assert len(response.chunks) == 2
+
+
+@pytest.mark.asyncio
+async def test_bm25_fallback_to_simple_search(milvus_index, sample_chunks, sample_embeddings, mock_milvus_client):
+    """Test that when BM25 search fails, the system falls back to simple text search."""
+    mock_milvus_client.has_collection.return_value = True
+    await milvus_index.add_chunks(sample_chunks, sample_embeddings)
+
+    # Force BM25 search to fail
+    mock_milvus_client.search.side_effect = Exception("BM25 search not available")
+
+    # Mock simple text search results
+    mock_milvus_client.query.return_value = [
+        {
+            "chunk_id": "chunk1",
+            "chunk_content": {"content": "Python programming language", "metadata": {"document_id": "doc1"}},
+        },
+        {
+            "chunk_id": "chunk2",
+            "chunk_content": {"content": "Machine learning algorithms", "metadata": {"document_id": "doc2"}},
+        },
+    ]
+
+    # Test keyword search that should fall back to simple text search
+    query_string = "Python"
+    response = await milvus_index.query_keyword(query_string=query_string, k=3, score_threshold=0.0)
+
+    # Verify response structure
+    assert isinstance(response, QueryChunksResponse)
+    assert len(response.chunks) > 0, "Fallback search should return results"
+
+    # Verify that simple text search was used (query method called instead of search)
+    mock_milvus_client.query.assert_called_once()
+    mock_milvus_client.search.assert_called_once()  # Called once but failed
+
+    # Verify the query uses parameterized filter with filter_params
+    query_call_args = mock_milvus_client.query.call_args
+    assert "filter" in query_call_args[1], "Query should include filter for text search"
+    assert "filter_params" in query_call_args[1], "Query should use parameterized filter"
+    assert query_call_args[1]["filter_params"]["content"] == "Python", "Filter params should contain the search term"
+
+    # Verify all returned chunks have score 1.0 (simple binary scoring)
+    assert all(score == 1.0 for score in response.scores), "Simple text search should use binary scoring"
+
+
+@pytest.mark.asyncio
+async def test_delete_collection(milvus_index, mock_milvus_client):
+    # Test collection deletion
+    mock_milvus_client.has_collection.return_value = True
+
+    await milvus_index.delete()
+
+    mock_milvus_client.drop_collection.assert_called_once_with(collection_name=milvus_index.collection_name)