Merge branch 'meta-llama:main' into dell-distro

llama_stack/providers/inline/safety/prompt_guard/prompt_guard.py

@@ -15,6 +15,7 @@ from llama_stack.apis.safety import (
     RunShieldResponse,
     Safety,
     SafetyViolation,
+    ShieldStore,
     ViolationLevel,
 )
 from llama_stack.apis.shields import Shield
@@ -32,6 +33,8 @@ PROMPT_GUARD_MODEL = "Prompt-Guard-86M"
 
 
 class PromptGuardSafetyImpl(Safety, ShieldsProtocolPrivate):
+    shield_store: ShieldStore
+
     def __init__(self, config: PromptGuardConfig, _deps) -> None:
         self.config = config
 
@@ -53,7 +56,7 @@ class PromptGuardSafetyImpl(Safety, ShieldsProtocolPrivate):
         self,
         shield_id: str,
         messages: list[Message],
-        params: dict[str, Any] = None,
+        params: dict[str, Any],
     ) -> RunShieldResponse:
         shield = await self.shield_store.get_shield(shield_id)
         if not shield:
@@ -61,6 +64,9 @@ class PromptGuardSafetyImpl(Safety, ShieldsProtocolPrivate):
 
         return await self.shield.run(messages)
 
+    async def run_moderation(self, input: str | list[str], model: str):
+        raise NotImplementedError("run_moderation not implemented for PromptGuard")
+
 
 class PromptGuardShield:
     def __init__(
@@ -117,8 +123,10 @@ class PromptGuardShield:
         elif self.config.guard_type == PromptGuardType.jailbreak.value and score_malicious > self.threshold:
             violation = SafetyViolation(
                 violation_level=ViolationLevel.ERROR,
-                violation_type=f"prompt_injection:malicious={score_malicious}",
-                violation_return_message="Sorry, I cannot do this.",
+                user_message="Sorry, I cannot do this.",
+                metadata={
+                    "violation_type": f"prompt_injection:malicious={score_malicious}",
+                },
             )
 
         return RunShieldResponse(violation=violation)

llama_stack/providers/inline/vector_io/faiss/faiss.py

@@ -174,7 +174,9 @@ class FaissIndex(EmbeddingIndex):
         k: int,
         score_threshold: float,
     ) -> QueryChunksResponse:
-        raise NotImplementedError("Keyword search is not supported in FAISS")
+        raise NotImplementedError(
+            "Keyword search is not supported - underlying DB FAISS does not support this search mode"
+        )
 
     async def query_hybrid(
         self,
@@ -185,7 +187,9 @@ class FaissIndex(EmbeddingIndex):
         reranker_type: str,
         reranker_params: dict[str, Any] | None = None,
     ) -> QueryChunksResponse:
-        raise NotImplementedError("Hybrid search is not supported in FAISS")
+        raise NotImplementedError(
+            "Hybrid search is not supported - underlying DB FAISS does not support this search mode"
+        )
 
 
 class FaissVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPrivate):

llama_stack/providers/registry/inference.py

@@ -213,6 +213,36 @@ def available_providers() -> list[ProviderSpec]:
                 description="Google Gemini inference provider for accessing Gemini models and Google's AI services.",
             ),
         ),
+        remote_provider_spec(
+            api=Api.inference,
+            adapter=AdapterSpec(
+                adapter_type="vertexai",
+                pip_packages=["litellm", "google-cloud-aiplatform"],
+                module="llama_stack.providers.remote.inference.vertexai",
+                config_class="llama_stack.providers.remote.inference.vertexai.VertexAIConfig",
+                provider_data_validator="llama_stack.providers.remote.inference.vertexai.config.VertexAIProviderDataValidator",
+                description="""Google Vertex AI inference provider enables you to use Google's Gemini models through Google Cloud's Vertex AI platform, providing several advantages:
+
+• Enterprise-grade security: Uses Google Cloud's security controls and IAM
+• Better integration: Seamless integration with other Google Cloud services
+• Advanced features: Access to additional Vertex AI features like model tuning and monitoring
+• Authentication: Uses Google Cloud Application Default Credentials (ADC) instead of API keys
+
+Configuration:
+- Set VERTEX_AI_PROJECT environment variable (required)
+- Set VERTEX_AI_LOCATION environment variable (optional, defaults to us-central1)
+- Use Google Cloud Application Default Credentials or service account key
+
+Authentication Setup:
+Option 1 (Recommended): gcloud auth application-default login
+Option 2: Set GOOGLE_APPLICATION_CREDENTIALS to service account key path
+
+Available Models:
+- vertex_ai/gemini-2.0-flash
+- vertex_ai/gemini-2.5-flash
+- vertex_ai/gemini-2.5-pro""",
+            ),
+        ),
         remote_provider_spec(
             api=Api.inference,
             adapter=AdapterSpec(

llama_stack/providers/registry/vector_io.py

@@ -45,6 +45,18 @@ That means you'll get fast and efficient vector retrieval.
 - Lightweight and easy to use
 - Fully integrated with Llama Stack
 - GPU support
+- **Vector search** - FAISS supports pure vector similarity search using embeddings
+
+## Search Modes
+
+**Supported:**
+- **Vector Search** (`mode="vector"`): Performs vector similarity search using embeddings
+
+**Not Supported:**
+- **Keyword Search** (`mode="keyword"`): Not supported by FAISS
+- **Hybrid Search** (`mode="hybrid"`): Not supported by FAISS
+
+> **Note**: FAISS is designed as a pure vector similarity search library. See the [FAISS GitHub repository](https://github.com/facebookresearch/faiss) for more details about FAISS's core functionality.
 
 ## Usage
 
@@ -535,6 +547,7 @@ That means you're not limited to storing vectors in memory or in a separate serv
 
 - Easy to use
 - Fully integrated with Llama Stack
+- Supports all search modes: vector, keyword, and hybrid search (both inline and remote configurations)
 
 ## Usage
 
@@ -625,6 +638,92 @@ vector_io:
 - **`client_pem_path`**: Path to the **client certificate** file (required for mTLS).
 - **`client_key_path`**: Path to the **client private key** file (required for mTLS).
 
+## Search Modes
+
+Milvus supports three different search modes for both inline and remote configurations:
+
+### Vector Search
+Vector search uses semantic similarity to find the most relevant chunks based on embedding vectors. This is the default search mode and works well for finding conceptually similar content.
+
+```python
+# Vector search example
+search_response = client.vector_stores.search(
+    vector_store_id=vector_store.id,
+    query="What is machine learning?",
+    search_mode="vector",
+    max_num_results=5,
+)
+```
+
+### Keyword Search
+Keyword search uses traditional text-based matching to find chunks containing specific terms or phrases. This is useful when you need exact term matches.
+
+```python
+# Keyword search example
+search_response = client.vector_stores.search(
+    vector_store_id=vector_store.id,
+    query="Python programming language",
+    search_mode="keyword",
+    max_num_results=5,
+)
+```
+
+### Hybrid Search
+Hybrid search combines both vector and keyword search methods to provide more comprehensive results. It leverages the strengths of both semantic similarity and exact term matching.
+
+#### Basic Hybrid Search
+```python
+# Basic hybrid search example (uses RRF ranker with default impact_factor=60.0)
+search_response = client.vector_stores.search(
+    vector_store_id=vector_store.id,
+    query="neural networks in Python",
+    search_mode="hybrid",
+    max_num_results=5,
+)
+```
+
+**Note**: The default `impact_factor` value of 60.0 was empirically determined to be optimal in the original RRF research paper: ["Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods"](https://plg.uwaterloo.ca/~gvcormac/cormacksigir09-rrf.pdf) (Cormack et al., 2009).
+
+#### Hybrid Search with RRF (Reciprocal Rank Fusion) Ranker
+RRF combines rankings from vector and keyword search by using reciprocal ranks. The impact factor controls how much weight is given to higher-ranked results.
+
+```python
+# Hybrid search with custom RRF parameters
+search_response = client.vector_stores.search(
+    vector_store_id=vector_store.id,
+    query="neural networks in Python",
+    search_mode="hybrid",
+    max_num_results=5,
+    ranking_options={
+        "ranker": {
+            "type": "rrf",
+            "impact_factor": 100.0,  # Higher values give more weight to top-ranked results
+        }
+    },
+)
+```
+
+#### Hybrid Search with Weighted Ranker
+Weighted ranker linearly combines normalized scores from vector and keyword search. The alpha parameter controls the balance between the two search methods.
+
+```python
+# Hybrid search with weighted ranker
+search_response = client.vector_stores.search(
+    vector_store_id=vector_store.id,
+    query="neural networks in Python",
+    search_mode="hybrid",
+    max_num_results=5,
+    ranking_options={
+        "ranker": {
+            "type": "weighted",
+            "alpha": 0.7,  # 70% vector search, 30% keyword search
+        }
+    },
+)
+```
+
+For detailed documentation on RRF and Weighted rankers, please refer to the [Milvus Reranking Guide](https://milvus.io/docs/reranking.md).
+
 ## Documentation
 See the [Milvus documentation](https://milvus.io/docs/install-overview.md) for more details about Milvus in general.
 

llama_stack/providers/remote/inference/gemini/models.py

@@ -13,7 +13,9 @@ LLM_MODEL_IDS = [
     "gemini-1.5-flash",
     "gemini-1.5-pro",
     "gemini-2.0-flash",
+    "gemini-2.0-flash-lite",
     "gemini-2.5-flash",
+    "gemini-2.5-flash-lite",
     "gemini-2.5-pro",
 ]
 

llama_stack/providers/remote/inference/ollama/ollama.py

@@ -457,9 +457,6 @@ class OllamaInferenceAdapter(
         user: str | None = None,
     ) -> OpenAIEmbeddingsResponse:
         model_obj = await self._get_model(model)
-        if model_obj.model_type != ModelType.embedding:
-            raise ValueError(f"Model {model} is not an embedding model")
-
         if model_obj.provider_resource_id is None:
             raise ValueError(f"Model {model} has no provider_resource_id set")
 

llama_stack/providers/remote/inference/vertexai/__init__.py

@@ -0,0 +1,15 @@
+# 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 .config import VertexAIConfig
+
+
+async def get_adapter_impl(config: VertexAIConfig, _deps):
+    from .vertexai import VertexAIInferenceAdapter
+
+    impl = VertexAIInferenceAdapter(config)
+    await impl.initialize()
+    return impl

llama_stack/providers/remote/inference/vertexai/config.py

@@ -0,0 +1,45 @@
+# 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 typing import Any
+
+from pydantic import BaseModel, Field
+
+from llama_stack.schema_utils import json_schema_type
+
+
+class VertexAIProviderDataValidator(BaseModel):
+    vertex_project: str | None = Field(
+        default=None,
+        description="Google Cloud project ID for Vertex AI",
+    )
+    vertex_location: str | None = Field(
+        default=None,
+        description="Google Cloud location for Vertex AI (e.g., us-central1)",
+    )
+
+
+@json_schema_type
+class VertexAIConfig(BaseModel):
+    project: str = Field(
+        description="Google Cloud project ID for Vertex AI",
+    )
+    location: str = Field(
+        default="us-central1",
+        description="Google Cloud location for Vertex AI",
+    )
+
+    @classmethod
+    def sample_run_config(
+        cls,
+        project: str = "${env.VERTEX_AI_PROJECT:=}",
+        location: str = "${env.VERTEX_AI_LOCATION:=us-central1}",
+        **kwargs,
+    ) -> dict[str, Any]:
+        return {
+            "project": project,
+            "location": location,
+        }

llama_stack/providers/remote/inference/vertexai/models.py

@@ -0,0 +1,20 @@
+# 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 llama_stack.providers.utils.inference.model_registry import (
+    ProviderModelEntry,
+)
+
+# Vertex AI model IDs with vertex_ai/ prefix as required by litellm
+LLM_MODEL_IDS = [
+    "vertex_ai/gemini-2.0-flash",
+    "vertex_ai/gemini-2.5-flash",
+    "vertex_ai/gemini-2.5-pro",
+]
+
+SAFETY_MODELS_ENTRIES = list[ProviderModelEntry]()
+
+MODEL_ENTRIES = [ProviderModelEntry(provider_model_id=m) for m in LLM_MODEL_IDS] + SAFETY_MODELS_ENTRIES

llama_stack/providers/remote/inference/vertexai/vertexai.py

@@ -0,0 +1,52 @@
+# 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 typing import Any
+
+from llama_stack.apis.inference import ChatCompletionRequest
+from llama_stack.providers.utils.inference.litellm_openai_mixin import (
+    LiteLLMOpenAIMixin,
+)
+
+from .config import VertexAIConfig
+from .models import MODEL_ENTRIES
+
+
+class VertexAIInferenceAdapter(LiteLLMOpenAIMixin):
+    def __init__(self, config: VertexAIConfig) -> None:
+        LiteLLMOpenAIMixin.__init__(
+            self,
+            MODEL_ENTRIES,
+            litellm_provider_name="vertex_ai",
+            api_key_from_config=None,  # Vertex AI uses ADC, not API keys
+            provider_data_api_key_field="vertex_project",  # Use project for validation
+        )
+        self.config = config
+
+    def get_api_key(self) -> str:
+        # Vertex AI doesn't use API keys, it uses Application Default Credentials
+        # Return empty string to let litellm handle authentication via ADC
+        return ""
+
+    async def _get_params(self, request: ChatCompletionRequest) -> dict[str, Any]:
+        # Get base parameters from parent
+        params = await super()._get_params(request)
+
+        # Add Vertex AI specific parameters
+        provider_data = self.get_request_provider_data()
+        if provider_data:
+            if getattr(provider_data, "vertex_project", None):
+                params["vertex_project"] = provider_data.vertex_project
+            if getattr(provider_data, "vertex_location", None):
+                params["vertex_location"] = provider_data.vertex_location
+        else:
+            params["vertex_project"] = self.config.project
+            params["vertex_location"] = self.config.location
+
+        # Remove api_key since Vertex AI uses ADC
+        params.pop("api_key", None)
+
+        return params

llama_stack/providers/utils/inference/openai_compat.py

@@ -70,7 +70,7 @@ from openai.types.chat.chat_completion_chunk import (
 from openai.types.chat.chat_completion_content_part_image_param import (
     ImageURL as OpenAIImageURL,
 )
-from openai.types.chat.chat_completion_message_tool_call_param import (
+from openai.types.chat.chat_completion_message_tool_call import (
     Function as OpenAIFunction,
 )
 from pydantic import BaseModel

llama_stack/providers/utils/telemetry/tracing.py

@@ -9,7 +9,9 @@ import contextvars
 import logging
 import queue
 import random
+import sys
 import threading
+import time
 from collections.abc import Callable
 from datetime import UTC, datetime
 from functools import wraps
@@ -30,6 +32,16 @@ from llama_stack.providers.utils.telemetry.trace_protocol import serialize_value
 
 logger = get_logger(__name__, category="core")
 
+# Fallback logger that does NOT propagate to TelemetryHandler to avoid recursion
+_fallback_logger = logging.getLogger("llama_stack.telemetry.background")
+if not _fallback_logger.handlers:
+    _fallback_logger.propagate = False
+    _fallback_logger.setLevel(logging.ERROR)
+    _fallback_handler = logging.StreamHandler(sys.stderr)
+    _fallback_handler.setLevel(logging.ERROR)
+    _fallback_handler.setFormatter(logging.Formatter("%(asctime)s [%(levelname)s] %(name)s: %(message)s"))
+    _fallback_logger.addHandler(_fallback_handler)
+
 
 INVALID_SPAN_ID = 0x0000000000000000
 INVALID_TRACE_ID = 0x00000000000000000000000000000000
@@ -79,19 +91,32 @@ def generate_trace_id() -> str:
 CURRENT_TRACE_CONTEXT = contextvars.ContextVar("trace_context", default=None)
 BACKGROUND_LOGGER = None
 
+LOG_QUEUE_FULL_LOG_INTERVAL_SECONDS = 60.0
+
 
 class BackgroundLogger:
     def __init__(self, api: Telemetry, capacity: int = 100000):
         self.api = api
-        self.log_queue = queue.Queue(maxsize=capacity)
+        self.log_queue: queue.Queue[Any] = queue.Queue(maxsize=capacity)
         self.worker_thread = threading.Thread(target=self._process_logs, daemon=True)
         self.worker_thread.start()
+        self._last_queue_full_log_time: float = 0.0
+        self._dropped_since_last_notice: int = 0
 
     def log_event(self, event):
         try:
             self.log_queue.put_nowait(event)
         except queue.Full:
-            logger.error("Log queue is full, dropping event")
+            # Aggregate drops and emit at most once per interval via fallback logger
+            self._dropped_since_last_notice += 1
+            current_time = time.time()
+            if current_time - self._last_queue_full_log_time >= LOG_QUEUE_FULL_LOG_INTERVAL_SECONDS:
+                _fallback_logger.error(
+                    "Log queue is full; dropped %d events since last notice",
+                    self._dropped_since_last_notice,
+                )
+                self._last_queue_full_log_time = current_time
+                self._dropped_since_last_notice = 0
 
     def _process_logs(self):
         while True: