llama-stack 0.3.5__py3-none-any.whl → 0.4.1__py3-none-any.whl

llama_stack/providers/utils/responses/responses_store.py

@@ -4,25 +4,22 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from llama_stack.apis.agents import (
-    Order,
-)
-from llama_stack.apis.agents.openai_responses import (
+from llama_stack.core.datatypes import AccessRule
+from llama_stack.core.storage.datatypes import ResponsesStoreReference, SqlStoreReference
+from llama_stack.core.storage.sqlstore.authorized_sqlstore import AuthorizedSqlStore
+from llama_stack.core.storage.sqlstore.sqlstore import sqlstore_impl
+from llama_stack.log import get_logger
+from llama_stack_api import (
     ListOpenAIResponseInputItem,
     ListOpenAIResponseObject,
     OpenAIDeleteResponseObject,
+    OpenAIMessageParam,
     OpenAIResponseInput,
     OpenAIResponseObject,
     OpenAIResponseObjectWithInput,
+    Order,
 )
-from llama_stack.apis.inference import OpenAIMessageParam
-from llama_stack.core.datatypes import AccessRule
-from llama_stack.core.storage.datatypes import ResponsesStoreReference, SqlStoreReference
-from llama_stack.log import get_logger
-
-from ..sqlstore.api import ColumnDefinition, ColumnType
-from ..sqlstore.authorized_sqlstore import AuthorizedSqlStore
-from ..sqlstore.sqlstore import sqlstore_impl
+from llama_stack_api.internal.sqlstore import ColumnDefinition, ColumnType
 
 logger = get_logger(name=__name__, category="openai_responses")
 
@@ -252,19 +249,12 @@ class ResponsesStore:
         # Serialize messages to dict format for JSON storage
         messages_data = [msg.model_dump() for msg in messages]
 
-        # Upsert: try insert first, update if exists
-        try:
-            await self.sql_store.insert(
-                table="conversation_messages",
-                data={"conversation_id": conversation_id, "messages": messages_data},
-            )
-        except Exception:
-            # If insert fails due to ID conflict, update existing record
-            await self.sql_store.update(
-                table="conversation_messages",
-                data={"messages": messages_data},
-                where={"conversation_id": conversation_id},
-            )
+        await self.sql_store.upsert(
+            table="conversation_messages",
+            data={"conversation_id": conversation_id, "messages": messages_data},
+            conflict_columns=["conversation_id"],
+            update_columns=["messages"],
+        )
 
         logger.debug(f"Stored {len(messages)} messages for conversation {conversation_id}")
 

llama_stack/providers/utils/scoring/aggregation_utils.py

@@ -6,8 +6,7 @@
 import statistics
 from typing import Any
 
-from llama_stack.apis.scoring import ScoringResultRow
-from llama_stack.apis.scoring_functions import AggregationFunctionType
+from llama_stack_api import AggregationFunctionType, ScoringResultRow
 
 
 def aggregate_accuracy(scoring_results: list[ScoringResultRow]) -> dict[str, Any]:

llama_stack/providers/utils/scoring/base_scoring_fn.py

@@ -6,9 +6,8 @@
 from abc import ABC, abstractmethod
 from typing import Any
 
-from llama_stack.apis.scoring import ScoringFnParams, ScoringResultRow
-from llama_stack.apis.scoring_functions import ScoringFn
 from llama_stack.providers.utils.scoring.aggregation_utils import aggregate_metrics
+from llama_stack_api import ScoringFn, ScoringFnParams, ScoringResultRow
 
 
 class BaseScoringFn(ABC):

llama_stack/providers/utils/tools/mcp.py

@@ -15,18 +15,55 @@ from mcp import types as mcp_types
 from mcp.client.sse import sse_client
 from mcp.client.streamable_http import streamablehttp_client
 
-from llama_stack.apis.common.content_types import ImageContentItem, InterleavedContentItem, TextContentItem
-from llama_stack.apis.tools import (
+from llama_stack.core.datatypes import AuthenticationRequiredError
+from llama_stack.log import get_logger
+from llama_stack.providers.utils.tools.ttl_dict import TTLDict
+from llama_stack_api import (
+    ImageContentItem,
+    InterleavedContentItem,
     ListToolDefsResponse,
+    TextContentItem,
     ToolDef,
     ToolInvocationResult,
+    _URLOrData,
 )
-from llama_stack.core.datatypes import AuthenticationRequiredError
-from llama_stack.log import get_logger
-from llama_stack.providers.utils.tools.ttl_dict import TTLDict
 
 logger = get_logger(__name__, category="tools")
 
+
+def prepare_mcp_headers(base_headers: dict[str, str] | None, authorization: str | None) -> dict[str, str]:
+    """
+    Prepare headers for MCP requests with authorization support.
+
+    Args:
+        base_headers: Base headers dictionary (can be None)
+        authorization: OAuth access token (without "Bearer " prefix)
+
+    Returns:
+        Headers dictionary with Authorization header if token provided
+
+    Raises:
+        ValueError: If Authorization header is specified in the headers dict (security risk)
+    """
+    headers = dict(base_headers or {})
+
+    # Security check: reject any Authorization header in the headers dict
+    # Users must use the authorization parameter instead to avoid security risks
+    existing_keys_lower = {k.lower() for k in headers.keys()}
+    if "authorization" in existing_keys_lower:
+        raise ValueError(
+            "For security reasons, Authorization header cannot be passed via 'headers'. "
+            "Please use the 'authorization' parameter instead."
+        )
+
+    # Add Authorization header if token provided
+    if authorization:
+        # OAuth access token - add "Bearer " prefix
+        headers["Authorization"] = f"Bearer {authorization}"
+
+    return headers
+
+
 protocol_cache = TTLDict(ttl_seconds=3600)
 
 
@@ -49,7 +86,10 @@ async def client_wrapper(endpoint: str, headers: dict[str, str]) -> AsyncGenerat
         try:
             client = streamablehttp_client
             if strategy == MCPProtol.SSE:
-                client = sse_client
+                # sse_client and streamablehttp_client have different signatures, but both
+                # are called the same way here, so we cast to Any to avoid type errors
+                client = cast(Any, sse_client)
+
             async with client(endpoint, headers=headers) as client_streams:
                 async with ClientSession(read_stream=client_streams[0], write_stream=client_streams[1]) as session:
                     await session.initialize()
@@ -107,9 +147,29 @@ async def client_wrapper(endpoint: str, headers: dict[str, str]) -> AsyncGenerat
                 raise
 
 
-async def list_mcp_tools(endpoint: str, headers: dict[str, str]) -> ListToolDefsResponse:
+async def list_mcp_tools(
+    endpoint: str,
+    headers: dict[str, str] | None = None,
+    authorization: str | None = None,
+) -> ListToolDefsResponse:
+    """List tools available from an MCP server.
+
+    Args:
+        endpoint: MCP server endpoint URL
+        headers: Optional base headers to include
+        authorization: Optional OAuth access token (just the token, not "Bearer <token>")
+
+    Returns:
+        List of tool definitions from the MCP server
+
+    Raises:
+        ValueError: If Authorization is found in the headers parameter
+    """
+    # Prepare headers with authorization handling
+    final_headers = prepare_mcp_headers(headers, authorization)
+
     tools = []
-    async with client_wrapper(endpoint, headers) as session:
+    async with client_wrapper(endpoint, final_headers) as session:
         tools_result = await session.list_tools()
         for tool in tools_result.tools:
             tools.append(
@@ -127,9 +187,31 @@ async def list_mcp_tools(endpoint: str, headers: dict[str, str]) -> ListToolDefs
 
 
 async def invoke_mcp_tool(
-    endpoint: str, headers: dict[str, str], tool_name: str, kwargs: dict[str, Any]
+    endpoint: str,
+    tool_name: str,
+    kwargs: dict[str, Any],
+    headers: dict[str, str] | None = None,
+    authorization: str | None = None,
 ) -> ToolInvocationResult:
-    async with client_wrapper(endpoint, headers) as session:
+    """Invoke an MCP tool with the given arguments.
+
+    Args:
+        endpoint: MCP server endpoint URL
+        tool_name: Name of the tool to invoke
+        kwargs: Tool invocation arguments
+        headers: Optional base headers to include
+        authorization: Optional OAuth access token (just the token, not "Bearer <token>")
+
+    Returns:
+        Tool invocation result with content and error information
+
+    Raises:
+        ValueError: If Authorization header is found in the headers parameter
+    """
+    # Prepare headers with authorization handling
+    final_headers = prepare_mcp_headers(headers, authorization)
+
+    async with client_wrapper(endpoint, final_headers) as session:
         result = await session.call_tool(tool_name, kwargs)
 
         content: list[InterleavedContentItem] = []
@@ -137,7 +219,7 @@ async def invoke_mcp_tool(
             if isinstance(item, mcp_types.TextContent):
                 content.append(TextContentItem(text=item.text))
             elif isinstance(item, mcp_types.ImageContent):
-                content.append(ImageContentItem(image=item.data))
+                content.append(ImageContentItem(image=_URLOrData(data=item.data)))
             elif isinstance(item, mcp_types.EmbeddedResource):
                 logger.warning(f"EmbeddedResource is not supported: {item}")
             else:

llama_stack/providers/utils/vector_io/__init__.py

@@ -3,3 +3,19 @@
 #
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
+
+from .vector_utils import (
+    WeightedInMemoryAggregator,
+    generate_chunk_id,
+    load_embedded_chunk_with_backward_compat,
+    proper_case,
+    sanitize_collection_name,
+)
+
+__all__ = [
+    "WeightedInMemoryAggregator",
+    "generate_chunk_id",
+    "load_embedded_chunk_with_backward_compat",
+    "proper_case",
+    "sanitize_collection_name",
+]

llama_stack/providers/utils/vector_io/vector_utils.py

@@ -7,6 +7,9 @@
 import hashlib
 import re
 import uuid
+from typing import Any
+
+from llama_stack_api import EmbeddedChunk
 
 
 def generate_chunk_id(document_id: str, chunk_text: str, chunk_window: str | None = None) -> str:
@@ -154,3 +157,36 @@ class WeightedInMemoryAggregator:
             # Default to RRF for None, RRF, or any unknown types
             impact_factor = reranker_params.get("impact_factor", 60.0)
             return WeightedInMemoryAggregator.rrf_rerank(vector_scores, keyword_scores, impact_factor)
+
+
+def load_embedded_chunk_with_backward_compat(
+    chunk_data: dict[str, Any],
+) -> EmbeddedChunk:
+    """
+    Load EmbeddedChunk data with backward compatibility for legacy field locations.
+
+    Handles migration from old format where embedding_model and embedding_dimension
+    were stored in chunk_metadata to current top-level format.
+
+    Args:
+        chunk_data: Dictionary containing chunk data to load
+
+    Returns:
+        EmbeddedChunk object with migrated data
+    """
+    # Migrate old data: extract embedding_model/embedding_dimension from chunk_metadata if missing
+    if "embedding_model" not in chunk_data:
+        chunk_metadata = chunk_data.get("chunk_metadata", {})
+        chunk_data["embedding_model"] = chunk_metadata.get("chunk_embedding_model", "unknown")
+
+    if "embedding_dimension" not in chunk_data:
+        chunk_metadata = chunk_data.get("chunk_metadata", {})
+        chunk_data["embedding_dimension"] = chunk_metadata.get(
+            "chunk_embedding_dimension", len(chunk_data.get("embedding", []))
+        )
+
+    # Ensure embedding field exists (required by EmbeddedChunk)
+    if "embedding" not in chunk_data:
+        chunk_data["embedding"] = []
+
+    return EmbeddedChunk(**chunk_data)

llama_stack/telemetry/constants.py

@@ -0,0 +1,27 @@
+# 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.
+
+"""
+This file contains constants used for naming data captured for telemetry.
+
+This is used to ensure that the data captured for telemetry is consistent and can be used to
+identify and correlate data. If custom telemetry data is added to llama stack, please add
+constants for it here.
+"""
+
+llama_stack_prefix = "llama_stack"
+
+# Safety Attributes
+RUN_SHIELD_OPERATION_NAME = "run_shield"
+
+SAFETY_REQUEST_PREFIX = f"{llama_stack_prefix}.safety.request"
+SAFETY_REQUEST_SHIELD_ID_ATTRIBUTE = f"{SAFETY_REQUEST_PREFIX}.shield_id"
+SAFETY_REQUEST_MESSAGES_ATTRIBUTE = f"{SAFETY_REQUEST_PREFIX}.messages"
+
+SAFETY_RESPONSE_PREFIX = f"{llama_stack_prefix}.safety.response"
+SAFETY_RESPONSE_METADATA_ATTRIBUTE = f"{SAFETY_RESPONSE_PREFIX}.metadata"
+SAFETY_RESPONSE_VIOLATION_LEVEL_ATTRIBUTE = f"{SAFETY_RESPONSE_PREFIX}.violation.level"
+SAFETY_RESPONSE_USER_MESSAGE_ATTRIBUTE = f"{SAFETY_RESPONSE_PREFIX}.violation.user_message"

llama_stack/telemetry/helpers.py

@@ -0,0 +1,43 @@
+# 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.
+
+import json
+
+from opentelemetry import trace
+
+from llama_stack_api import OpenAIMessageParam, RunShieldResponse
+
+from .constants import (
+    RUN_SHIELD_OPERATION_NAME,
+    SAFETY_REQUEST_MESSAGES_ATTRIBUTE,
+    SAFETY_REQUEST_SHIELD_ID_ATTRIBUTE,
+    SAFETY_RESPONSE_METADATA_ATTRIBUTE,
+    SAFETY_RESPONSE_USER_MESSAGE_ATTRIBUTE,
+    SAFETY_RESPONSE_VIOLATION_LEVEL_ATTRIBUTE,
+)
+
+
+def safety_span_name(shield_id: str) -> str:
+    return f"{RUN_SHIELD_OPERATION_NAME} {shield_id}"
+
+
+# TODO: Consider using Wrapt to automatically instrument code
+# This is the industry standard way to package automatically instrumentation in python.
+def safety_request_span_attributes(
+    shield_id: str, messages: list[OpenAIMessageParam], response: RunShieldResponse
+) -> None:
+    span = trace.get_current_span()
+    span.set_attribute(SAFETY_REQUEST_SHIELD_ID_ATTRIBUTE, shield_id)
+    messages_json = json.dumps([msg.model_dump() for msg in messages])
+    span.set_attribute(SAFETY_REQUEST_MESSAGES_ATTRIBUTE, messages_json)
+
+    if response.violation:
+        if response.violation.metadata:
+            metadata_json = json.dumps(response.violation.metadata)
+            span.set_attribute(SAFETY_RESPONSE_METADATA_ATTRIBUTE, metadata_json)
+        if response.violation.user_message:
+            span.set_attribute(SAFETY_RESPONSE_USER_MESSAGE_ATTRIBUTE, response.violation.user_message)
+        span.set_attribute(SAFETY_RESPONSE_VIOLATION_LEVEL_ATTRIBUTE, response.violation.violation_level.value)

llama_stack/testing/api_recorder.py

@@ -40,10 +40,12 @@ from openai.types.completion_choice import CompletionChoice
 from llama_stack.core.testing_context import get_test_context, is_debug_mode
 
 # update the "finish_reason" field, since its type definition is wrong (no None is accepted)
-CompletionChoice.model_fields["finish_reason"].annotation = Literal["stop", "length", "content_filter"] | None
+CompletionChoice.model_fields["finish_reason"].annotation = cast(
+    type[Any] | None, Literal["stop", "length", "content_filter"] | None
+)
 CompletionChoice.model_rebuild()
 
-REPO_ROOT = Path(__file__).parent.parent.parent
+REPO_ROOT = Path(__file__).parent.parent.parent.parent
 DEFAULT_STORAGE_DIR = REPO_ROOT / "tests/integration/common"
 
 
@@ -154,7 +156,7 @@ def normalize_inference_request(method: str, url: str, headers: dict[str, Any],
     }
 
     # Include test_id for isolation, except for shared infrastructure endpoints
-    if parsed.path not in ("/api/tags", "/v1/models"):
+    if parsed.path not in ("/api/tags", "/v1/models", "/v1/openai/v1/models"):
         normalized["test_id"] = test_id
 
     normalized_json = json.dumps(normalized, sort_keys=True)
@@ -428,7 +430,7 @@ class ResponseStorage:
 
         # For model-list endpoints, include digest in filename to distinguish different model sets
         endpoint = request.get("endpoint")
-        if endpoint in ("/api/tags", "/v1/models"):
+        if endpoint in ("/api/tags", "/v1/models", "/v1/openai/v1/models"):
             digest = _model_identifiers_digest(endpoint, response)
             response_file = f"models-{request_hash}-{digest}.json"
 
@@ -552,13 +554,14 @@ def _model_identifiers_digest(endpoint: str, response: dict[str, Any]) -> str:
         Supported endpoints:
         - '/api/tags' (Ollama): response body has 'models': [ { name/model/digest/id/... }, ... ]
         - '/v1/models' (OpenAI): response body is: [ { id: ... }, ... ]
+        - '/v1/openai/v1/models' (OpenAI): response body is: [ { id: ... }, ... ]
         Returns a list of unique identifiers or None if structure doesn't match.
         """
         if "models" in response["body"]:
             # ollama
             items = response["body"]["models"]
         else:
-            # openai
+            # openai or openai-style endpoints
             items = response["body"]
         idents = [m.model if endpoint == "/api/tags" else m.id for m in items]
         return sorted(set(idents))
@@ -579,7 +582,7 @@ def _combine_model_list_responses(endpoint: str, records: list[dict[str, Any]])
     seen: dict[str, dict[str, Any]] = {}
     for rec in records:
         body = rec["response"]["body"]
-        if endpoint == "/v1/models":
+        if endpoint in ("/v1/models", "/v1/openai/v1/models"):
             for m in body:
                 key = m.id
                 seen[key] = m
@@ -597,19 +600,23 @@ def _combine_model_list_responses(endpoint: str, records: list[dict[str, Any]])
     if endpoint == "/api/tags":
         from ollama import ListResponse
 
-        body = ListResponse(models=ordered)
+        # Both cast(Any, ...) and type: ignore are needed here:
+        # - cast(Any, ...) attempts to bypass type checking on the argument
+        # - type: ignore is still needed because mypy checks the call site independently
+        #   and reports arg-type mismatch even after casting
+        body = ListResponse(models=cast(Any, ordered))  # type: ignore[arg-type]
     return {"request": canonical_req, "response": {"body": body, "is_streaming": False}}
 
 
 async def _patched_tool_invoke_method(
-    original_method, provider_name: str, self, tool_name: str, kwargs: dict[str, Any]
+    original_method, provider_name: str, self, tool_name: str, kwargs: dict[str, Any], authorization: str | None = None
 ):
     """Patched version of tool runtime invoke_tool method for recording/replay."""
     global _current_mode, _current_storage
 
     if _current_mode == APIRecordingMode.LIVE or _current_storage is None:
         # Normal operation
-        return await original_method(self, tool_name, kwargs)
+        return await original_method(self, tool_name, kwargs, authorization=authorization)
 
     request_hash = normalize_tool_request(provider_name, tool_name, kwargs)
 
@@ -627,7 +634,7 @@ async def _patched_tool_invoke_method(
 
     if _current_mode in (APIRecordingMode.RECORD, APIRecordingMode.RECORD_IF_MISSING):
         # Make the tool call and record it
-        result = await original_method(self, tool_name, kwargs)
+        result = await original_method(self, tool_name, kwargs, authorization=authorization)
 
         request_data = {
             "test_id": get_test_context(),
@@ -659,7 +666,7 @@ async def _patched_inference_method(original_method, self, client_type, endpoint
         logger.info(f"  Test context: {get_test_context()}")
 
     if mode == APIRecordingMode.LIVE or storage is None:
-        if endpoint == "/v1/models":
+        if endpoint in ("/v1/models", "/v1/openai/v1/models"):
             return original_method(self, *args, **kwargs)
         else:
             return await original_method(self, *args, **kwargs)
@@ -693,7 +700,7 @@ async def _patched_inference_method(original_method, self, client_type, endpoint
     recording = None
     if mode == APIRecordingMode.REPLAY or mode == APIRecordingMode.RECORD_IF_MISSING:
         # Special handling for model-list endpoints: merge all recordings with this hash
-        if endpoint in ("/api/tags", "/v1/models"):
+        if endpoint in ("/api/tags", "/v1/models", "/v1/openai/v1/models"):
             records = storage._model_list_responses(request_hash)
             recording = _combine_model_list_responses(endpoint, records)
         else:
@@ -733,13 +740,13 @@ async def _patched_inference_method(original_method, self, client_type, endpoint
             )
 
     if mode == APIRecordingMode.RECORD or (mode == APIRecordingMode.RECORD_IF_MISSING and not recording):
-        if endpoint == "/v1/models":
+        if endpoint in ("/v1/models", "/v1/openai/v1/models"):
             response = original_method(self, *args, **kwargs)
         else:
             response = await original_method(self, *args, **kwargs)
 
         # we want to store the result of the iterator, not the iterator itself
-        if endpoint == "/v1/models":
+        if endpoint in ("/v1/models", "/v1/openai/v1/models"):
             response = [m async for m in response]
 
         request_data = {
@@ -878,9 +885,11 @@ def patch_inference_clients():
     OllamaAsyncClient.list = patched_ollama_list
 
     # Create patched methods for tool runtimes
-    async def patched_tavily_invoke_tool(self, tool_name: str, kwargs: dict[str, Any]):
+    async def patched_tavily_invoke_tool(
+        self, tool_name: str, kwargs: dict[str, Any], authorization: str | None = None
+    ):
         return await _patched_tool_invoke_method(
-            _original_methods["tavily_invoke_tool"], "tavily", self, tool_name, kwargs
+            _original_methods["tavily_invoke_tool"], "tavily", self, tool_name, kwargs, authorization=authorization
         )
 
     # Apply tool runtime patches