phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-26 12:02:04 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge upstream/main and resolve conflicts

Browse source 
Resolved merge conflicts in:
- Documentation files: updated vector IO provider docs to include both kvstore fields and embedding model configuration
- Config files: merged kvstore requirements from upstream with embedding model fields
- Dependencies: updated to latest client versions while preserving llama-models dependency
- Regenerated lockfiles to ensure consistency

All embedding model configuration features preserved while incorporating upstream changes.
This commit is contained in:
skamenan7 2025-07-16 19:57:02 -04:00
commit 6634b21a76
92 changed files with 3069 additions and 2481 deletions
Download patch file Download diff file Expand all files Collapse all files

3
llama_stack/apis/vector_dbs/vector_dbs.py
View file

@ -19,6 +19,7 @@ class VectorDB(Resource):
embedding_model: str
embedding_dimension: int
vector_db_name: str | None = None
@property
def vector_db_id(self) -> str:
@ -70,6 +71,7 @@ class VectorDBs(Protocol):
embedding_model: str,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
vector_db_name: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorDB:
"""Register a vector database.
@ -78,6 +80,7 @@ class VectorDBs(Protocol):
:param embedding_model: The embedding model to use.
:param embedding_dimension: The dimension of the embedding model.
:param provider_id: The identifier of the provider.
:param vector_db_name: The name of the vector database.
:param provider_vector_db_id: The identifier of the vector database in the provider.
:returns: A VectorDB.
"""

2
llama_stack/apis/vector_io/vector_io.py
View file

@ -346,7 +346,6 @@ class VectorIO(Protocol):
embedding_model: str | None = None,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorStoreObject:
"""Creates a vector store.
@ -358,7 +357,6 @@ class VectorIO(Protocol):
:param embedding_model: The embedding model to use for this vector store.
:param embedding_dimension: The dimension of the embedding vectors (default: 384).
:param provider_id: The ID of the provider to use for this vector store.
:param provider_vector_db_id: The provider-specific vector database ID.
:returns: A VectorStoreObject representing the created vector store.
"""
...

8
llama_stack/distribution/configure.py
View file

@ -17,7 +17,7 @@ from llama_stack.distribution.distribution import (
builtin_automatically_routed_apis,
get_provider_registry,
)
from llama_stack.distribution.stack import replace_env_vars
from llama_stack.distribution.stack import cast_image_name_to_string, replace_env_vars
from llama_stack.distribution.utils.config_dirs import EXTERNAL_PROVIDERS_DIR
from llama_stack.distribution.utils.dynamic import instantiate_class_type
from llama_stack.distribution.utils.prompt_for_config import prompt_for_config
@ -164,7 +164,8 @@ def upgrade_from_routing_table(
def parse_and_maybe_upgrade_config(config_dict: dict[str, Any]) -> StackRunConfig:
version = config_dict.get("version", None)
if version == LLAMA_STACK_RUN_CONFIG_VERSION:
return StackRunConfig(**replace_env_vars(config_dict))
processed_config_dict = replace_env_vars(config_dict)
return StackRunConfig(**cast_image_name_to_string(processed_config_dict))
if "routing_table" in config_dict:
logger.info("Upgrading config...")
@ -175,4 +176,5 @@ def parse_and_maybe_upgrade_config(config_dict: dict[str, Any]) -> StackRunConfi
if not config_dict.get("external_providers_dir", None):
config_dict["external_providers_dir"] = EXTERNAL_PROVIDERS_DIR
return StackRunConfig(**replace_env_vars(config_dict))
processed_config_dict = replace_env_vars(config_dict)
return StackRunConfig(**cast_image_name_to_string(processed_config_dict))

2
llama_stack/distribution/resolver.py
View file

@ -200,7 +200,7 @@ def validate_and_prepare_providers(
specs = {}
for provider in providers:
if not provider.provider_id or provider.provider_id == "__disabled__":
logger.warning(f"Provider `{provider.provider_type}` for API `{api}` is disabled")
logger.debug(f"Provider `{provider.provider_type}` for API `{api}` is disabled")
continue
validate_provider(provider, api, provider_registry)

20
llama_stack/distribution/routers/vector_io.py
View file

@ -5,6 +5,7 @@
# the root directory of this source tree.
import asyncio
import uuid
from typing import Any
from llama_stack.apis.common.content_types import InterleavedContent
@ -105,6 +106,7 @@ class VectorIORouter(VectorIO):
embedding_model: str,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
vector_db_name: str | None = None,
provider_vector_db_id: str | None = None,
) -> None:
logger.debug(f"VectorIORouter.register_vector_db: {vector_db_id}, {embedding_model}")
@ -113,6 +115,7 @@ class VectorIORouter(VectorIO):
embedding_model,
embedding_dimension,
provider_id,
vector_db_name,
provider_vector_db_id,
)
@ -147,7 +150,6 @@ class VectorIORouter(VectorIO):
embedding_model: str | None = None,
embedding_dimension: int | None = None,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorStoreObject:
logger.debug(f"VectorIORouter.openai_create_vector_store: name={name}, provider_id={provider_id}")
@ -210,17 +212,17 @@ class VectorIORouter(VectorIO):
)
raise ValueError(f"Unable to determine embedding model for vector store '{name}': {e}") from e
vector_db_id = name
vector_db_id = f"vs_{uuid.uuid4()}"
registered_vector_db = await self.routing_table.register_vector_db(
vector_db_id,
embedding_model,
embedding_dimension,
provider_id,
provider_vector_db_id,
vector_db_id=vector_db_id,
embedding_model=embedding_model,
embedding_dimension=embedding_dimension,
provider_id=provider_id,
provider_vector_db_id=vector_db_id,
vector_db_name=name,
)
return await self.routing_table.get_provider_impl(registered_vector_db.identifier).openai_create_vector_store(
vector_db_id,
name=name,
file_ids=file_ids,
expires_after=expires_after,
chunking_strategy=chunking_strategy,

2
llama_stack/distribution/routing_tables/vector_dbs.py
View file

@ -36,6 +36,7 @@ class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
vector_db_name: str | None = None,
) -> VectorDB:
if provider_vector_db_id is None:
provider_vector_db_id = vector_db_id
@ -62,6 +63,7 @@ class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
"provider_resource_id": provider_vector_db_id,
"embedding_model": embedding_model,
"embedding_dimension": model.metadata["embedding_dimension"],
"vector_db_name": vector_db_name,
}
vector_db = TypeAdapter(VectorDBWithOwner).validate_python(vector_db_data)
await self.register_object(vector_db)

3
llama_stack/distribution/server/server.py
View file

@ -47,6 +47,7 @@ from llama_stack.distribution.server.routes import (
initialize_route_impls,
)
from llama_stack.distribution.stack import (
cast_image_name_to_string,
construct_stack,
replace_env_vars,
validate_env_pair,
@ -439,7 +440,7 @@ def main(args: argparse.Namespace | None = None):
logger.error(f"Error: {str(e)}")
sys.exit(1)
config = replace_env_vars(config_contents)
config = StackRunConfig(**config)
config = StackRunConfig(**cast_image_name_to_string(config))
# now that the logger is initialized, print the line about which type of config we are using.
logger.info(log_line)

7
llama_stack/distribution/stack.py
View file

@ -267,6 +267,13 @@ def _convert_string_to_proper_type(value: str) -> Any:
return value
def cast_image_name_to_string(config_dict: dict[str, Any]) -> dict[str, Any]:
"""Ensure that any value for a key 'image_name' in a config_dict is a string"""
if "image_name" in config_dict and config_dict["image_name"] is not None:
config_dict["image_name"] = str(config_dict["image_name"])
return config_dict
def validate_env_pair(env_pair: str) -> tuple[str, str]:
"""Validate and split an environment variable key-value pair."""
try:

3
llama_stack/providers/inline/files/localfs/files.py
View file

@ -51,6 +51,9 @@ class LocalfsFilesImpl(Files):
},
)
async def shutdown(self) -> None:
pass
def _generate_file_id(self) -> str:
"""Generate a unique file ID for OpenAI API."""
return f"file-{uuid.uuid4().hex}"

40
llama_stack/providers/inline/vector_io/faiss/faiss.py
View file

@ -181,8 +181,8 @@ class FaissVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPr
)
self.cache[vector_db.identifier] = index
# Load existing OpenAI vector stores using the mixin method
self.openai_vector_stores = await self._load_openai_vector_stores()
# Load existing OpenAI vector stores into the in-memory cache
await self.initialize_openai_vector_stores()
async def shutdown(self) -> None:
# Cleanup if needed
@ -261,42 +261,6 @@ class FaissVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPr
return await index.query_chunks(query, params)
# OpenAI Vector Store Mixin abstract method implementations
async def _save_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Save vector store metadata to kvstore."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
self.openai_vector_stores[store_id] = store_info
async def _load_openai_vector_stores(self) -> dict[str, dict[str, Any]]:
"""Load all vector store metadata from kvstore."""
assert self.kvstore is not None
start_key = OPENAI_VECTOR_STORES_PREFIX
end_key = f"{OPENAI_VECTOR_STORES_PREFIX}\xff"
stored_openai_stores = await self.kvstore.values_in_range(start_key, end_key)
stores = {}
for store_data in stored_openai_stores:
store_info = json.loads(store_data)
stores[store_info["id"]] = store_info
return stores
async def _update_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Update vector store metadata in kvstore."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
self.openai_vector_stores[store_id] = store_info
async def _delete_openai_vector_store_from_storage(self, store_id: str) -> None:
"""Delete vector store metadata from kvstore."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.delete(key)
if store_id in self.openai_vector_stores:
del self.openai_vector_stores[store_id]
async def _save_openai_vector_store_file(
self, store_id: str, file_id: str, file_info: dict[str, Any], file_contents: list[dict[str, Any]]
) -> None:

80
llama_stack/providers/inline/vector_io/sqlite_vec/sqlite_vec.py
View file

@ -7,6 +7,7 @@
import asyncio
import json
import logging
import re
import sqlite3
import struct
from typing import Any
@ -117,6 +118,10 @@ def _rrf_rerank(
return rrf_scores
def _make_sql_identifier(name: str) -> str:
return re.sub(r"[^a-zA-Z0-9_]", "_", name)
class SQLiteVecIndex(EmbeddingIndex):
"""
An index implementation that stores embeddings in a SQLite virtual table using sqlite-vec.
@ -130,9 +135,9 @@ class SQLiteVecIndex(EmbeddingIndex):
self.dimension = dimension
self.db_path = db_path
self.bank_id = bank_id
self.metadata_table = f"chunks_{bank_id}".replace("-", "_")
self.vector_table = f"vec_chunks_{bank_id}".replace("-", "_")
self.fts_table = f"fts_chunks_{bank_id}".replace("-", "_")
self.metadata_table = _make_sql_identifier(f"chunks_{bank_id}")
self.vector_table = _make_sql_identifier(f"vec_chunks_{bank_id}")
self.fts_table = _make_sql_identifier(f"fts_chunks_{bank_id}")
self.kvstore = kvstore
@classmethod
@ -148,14 +153,14 @@ class SQLiteVecIndex(EmbeddingIndex):
try:
# Create the table to store chunk metadata.
cur.execute(f"""
CREATE TABLE IF NOT EXISTS {self.metadata_table} (
CREATE TABLE IF NOT EXISTS [{self.metadata_table}] (
id TEXT PRIMARY KEY,
chunk TEXT
);
""")
# Create the virtual table for embeddings.
cur.execute(f"""
CREATE VIRTUAL TABLE IF NOT EXISTS {self.vector_table}
CREATE VIRTUAL TABLE IF NOT EXISTS [{self.vector_table}]
USING vec0(embedding FLOAT[{self.dimension}], id TEXT);
""")
connection.commit()
@ -163,7 +168,7 @@ class SQLiteVecIndex(EmbeddingIndex):
# based on query. Implementation of the change on client side will allow passing the search_mode option
# during initialization to make it easier to create the table that is required.
cur.execute(f"""
CREATE VIRTUAL TABLE IF NOT EXISTS {self.fts_table}
CREATE VIRTUAL TABLE IF NOT EXISTS [{self.fts_table}]
USING fts5(id, content);
""")
connection.commit()
@ -178,9 +183,9 @@ class SQLiteVecIndex(EmbeddingIndex):
connection = _create_sqlite_connection(self.db_path)
cur = connection.cursor()
try:
cur.execute(f"DROP TABLE IF EXISTS {self.metadata_table};")
cur.execute(f"DROP TABLE IF EXISTS {self.vector_table};")
cur.execute(f"DROP TABLE IF EXISTS {self.fts_table};")
cur.execute(f"DROP TABLE IF EXISTS [{self.metadata_table}];")
cur.execute(f"DROP TABLE IF EXISTS [{self.vector_table}];")
cur.execute(f"DROP TABLE IF EXISTS [{self.fts_table}];")
connection.commit()
finally:
cur.close()
@ -212,7 +217,7 @@ class SQLiteVecIndex(EmbeddingIndex):
metadata_data = [(chunk.chunk_id, chunk.model_dump_json()) for chunk in batch_chunks]
cur.executemany(
f"""
INSERT INTO {self.metadata_table} (id, chunk)
INSERT INTO [{self.metadata_table}] (id, chunk)
VALUES (?, ?)
ON CONFLICT(id) DO UPDATE SET chunk = excluded.chunk;
""",
@ -230,7 +235,7 @@ class SQLiteVecIndex(EmbeddingIndex):
for chunk, emb in zip(batch_chunks, batch_embeddings, strict=True)
]
cur.executemany(
f"INSERT INTO {self.vector_table} (id, embedding) VALUES (?, ?);",
f"INSERT INTO [{self.vector_table}] (id, embedding) VALUES (?, ?);",
embedding_data,
)
@ -238,13 +243,13 @@ class SQLiteVecIndex(EmbeddingIndex):
fts_data = [(chunk.chunk_id, chunk.content) for chunk in batch_chunks]
# DELETE existing entries with same IDs (FTS5 doesn't support ON CONFLICT)
cur.executemany(
f"DELETE FROM {self.fts_table} WHERE id = ?;",
f"DELETE FROM [{self.fts_table}] WHERE id = ?;",
[(row[0],) for row in fts_data],
)
# INSERT new entries
cur.executemany(
f"INSERT INTO {self.fts_table} (id, content) VALUES (?, ?);",
f"INSERT INTO [{self.fts_table}] (id, content) VALUES (?, ?);",
fts_data,
)
@ -280,8 +285,8 @@ class SQLiteVecIndex(EmbeddingIndex):
emb_blob = serialize_vector(emb_list)
query_sql = f"""
SELECT m.id, m.chunk, v.distance
FROM {self.vector_table} AS v
JOIN {self.metadata_table} AS m ON m.id = v.id
FROM [{self.vector_table}] AS v
JOIN [{self.metadata_table}] AS m ON m.id = v.id
WHERE v.embedding MATCH ? AND k = ?
ORDER BY v.distance;
"""
@ -322,9 +327,9 @@ class SQLiteVecIndex(EmbeddingIndex):
cur = connection.cursor()
try:
query_sql = f"""
SELECT DISTINCT m.id, m.chunk, bm25({self.fts_table}) AS score
FROM {self.fts_table} AS f
JOIN {self.metadata_table} AS m ON m.id = f.id
SELECT DISTINCT m.id, m.chunk, bm25([{self.fts_table}]) AS score
FROM [{self.fts_table}] AS f
JOIN [{self.metadata_table}] AS m ON m.id = f.id
WHERE f.content MATCH ?
ORDER BY score ASC
LIMIT ?;
@ -452,8 +457,8 @@ class SQLiteVecVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtoc
)
self.cache[vector_db.identifier] = VectorDBWithIndex(vector_db, index, self.inference_api)
# load any existing OpenAI vector stores
self.openai_vector_stores = await self._load_openai_vector_stores()
# Load existing OpenAI vector stores into the in-memory cache
await self.initialize_openai_vector_stores()
async def shutdown(self) -> None:
# nothing to do since we don't maintain a persistent connection
@ -501,41 +506,6 @@ class SQLiteVecVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtoc
await self.cache[vector_db_id].index.delete()
del self.cache[vector_db_id]
# OpenAI Vector Store Mixin abstract method implementations
async def _save_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Save vector store metadata to SQLite database."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
self.openai_vector_stores[store_id] = store_info
async def _load_openai_vector_stores(self) -> dict[str, dict[str, Any]]:
"""Load all vector store metadata from SQLite database."""
assert self.kvstore is not None
start_key = OPENAI_VECTOR_STORES_PREFIX
end_key = f"{OPENAI_VECTOR_STORES_PREFIX}\xff"
stored_openai_stores = await self.kvstore.values_in_range(start_key, end_key)
stores = {}
for store_data in stored_openai_stores:
store_info = json.loads(store_data)
stores[store_info["id"]] = store_info
return stores
async def _update_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Update vector store metadata in SQLite database."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
self.openai_vector_stores[store_id] = store_info
async def _delete_openai_vector_store_from_storage(self, store_id: str) -> None:
"""Delete vector store metadata from SQLite database."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.delete(key)
if store_id in self.openai_vector_stores:
del self.openai_vector_stores[store_id]
async def _save_openai_vector_store_file(
self, store_id: str, file_id: str, file_info: dict[str, Any], file_contents: list[dict[str, Any]]
) -> None:

37
llama_stack/providers/remote/inference/llama_openai_compat/llama.py
View file

@ -3,16 +3,17 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import logging
from llama_stack.providers.remote.inference.llama_openai_compat.config import (
LlamaCompatConfig,
)
from llama_stack.providers.utils.inference.litellm_openai_mixin import (
LiteLLMOpenAIMixin,
)
from llama_api_client import AsyncLlamaAPIClient, NotFoundError
from llama_stack.providers.remote.inference.llama_openai_compat.config import LlamaCompatConfig
from llama_stack.providers.utils.inference.litellm_openai_mixin import LiteLLMOpenAIMixin
from .models import MODEL_ENTRIES
logger = logging.getLogger(__name__)
class LlamaCompatInferenceAdapter(LiteLLMOpenAIMixin):
_config: LlamaCompatConfig
@ -27,8 +28,32 @@ class LlamaCompatInferenceAdapter(LiteLLMOpenAIMixin):
)
self.config = config
async def check_model_availability(self, model: str) -> bool:
"""
Check if a specific model is available from Llama API.
:param model: The model identifier to check.
:return: True if the model is available dynamically, False otherwise.
"""
try:
llama_api_client = self._get_llama_api_client()
retrieved_model = await llama_api_client.models.retrieve(model)
logger.info(f"Model {retrieved_model.id} is available from Llama API")
return True
except NotFoundError:
logger.error(f"Model {model} is not available from Llama API")
return False
except Exception as e:
logger.error(f"Failed to check model availability from Llama API: {e}")
return False
async def initialize(self):
await super().initialize()
async def shutdown(self):
await super().shutdown()
def _get_llama_api_client(self) -> AsyncLlamaAPIClient:
return AsyncLlamaAPIClient(api_key=self.get_api_key(), base_url=self.config.openai_compat_api_base)

47
llama_stack/providers/remote/inference/nvidia/nvidia.py
View file

@ -7,7 +7,6 @@
import logging
import warnings
from collections.abc import AsyncIterator
from functools import lru_cache
from typing import Any
from openai import APIConnectionError, AsyncOpenAI, BadRequestError
@ -93,41 +92,21 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
self._config = config
@lru_cache # noqa: B019
def _get_client(self, provider_model_id: str) -> AsyncOpenAI:
@property
def _client(self) -> AsyncOpenAI:
"""
For hosted models, https://integrate.api.nvidia.com/v1 is the primary base_url. However,
some models are hosted on different URLs. This function returns the appropriate client
for the given provider_model_id.
Returns an OpenAI client for the configured NVIDIA API endpoint.
This relies on lru_cache and self._default_client to avoid creating a new client for each request
or for each model that is hosted on https://integrate.api.nvidia.com/v1.
:param provider_model_id: The provider model ID
:return: An OpenAI client
"""
@lru_cache # noqa: B019
def _get_client_for_base_url(base_url: str) -> AsyncOpenAI:
"""
Maintain a single OpenAI client per base_url.
"""
return AsyncOpenAI(
base_url=base_url,
api_key=(self._config.api_key.get_secret_value() if self._config.api_key else "NO KEY"),
timeout=self._config.timeout,
)
special_model_urls = {
"meta/llama-3.2-11b-vision-instruct": "https://ai.api.nvidia.com/v1/gr/meta/llama-3.2-11b-vision-instruct",
"meta/llama-3.2-90b-vision-instruct": "https://ai.api.nvidia.com/v1/gr/meta/llama-3.2-90b-vision-instruct",
}
base_url = f"{self._config.url}/v1" if self._config.append_api_version else self._config.url
if _is_nvidia_hosted(self._config) and provider_model_id in special_model_urls:
base_url = special_model_urls[provider_model_id]
return _get_client_for_base_url(base_url)
return AsyncOpenAI(
base_url=base_url,
api_key=(self._config.api_key.get_secret_value() if self._config.api_key else "NO KEY"),
timeout=self._config.timeout,
)
async def _get_provider_model_id(self, model_id: str) -> str:
if not self.model_store:
@ -169,7 +148,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
response = await self._get_client(provider_model_id).completions.create(**request)
response = await self._client.completions.create(**request)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e
@ -222,7 +201,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
extra_body["input_type"] = task_type_options[task_type]
try:
response = await self._get_client(provider_model_id).embeddings.create(
response = await self._client.embeddings.create(
model=provider_model_id,
input=input,
extra_body=extra_body,
@ -283,7 +262,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
response = await self._get_client(provider_model_id).chat.completions.create(**request)
response = await self._client.chat.completions.create(**request)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e
@ -339,7 +318,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
return await self._get_client(provider_model_id).completions.create(**params)
return await self._client.completions.create(**params)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e
@ -398,7 +377,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
return await self._get_client(provider_model_id).chat.completions.create(**params)
return await self._client.chat.completions.create(**params)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e

25
llama_stack/providers/remote/inference/ollama/models.py
View file

@ -12,6 +12,19 @@ from llama_stack.providers.utils.inference.model_registry import (
build_model_entry,
)
SAFETY_MODELS_ENTRIES = [
# The Llama Guard models don't have their full fp16 versions
# so we are going to alias their default version to the canonical SKU
build_hf_repo_model_entry(
"llama-guard3:8b",
CoreModelId.llama_guard_3_8b.value,
),
build_hf_repo_model_entry(
"llama-guard3:1b",
CoreModelId.llama_guard_3_1b.value,
),
]
MODEL_ENTRIES = [
build_hf_repo_model_entry(
"llama3.1:8b-instruct-fp16",
@ -73,16 +86,6 @@ MODEL_ENTRIES = [
"llama3.3:70b",
CoreModelId.llama3_3_70b_instruct.value,
),
# The Llama Guard models don't have their full fp16 versions
# so we are going to alias their default version to the canonical SKU
build_hf_repo_model_entry(
"llama-guard3:8b",
CoreModelId.llama_guard_3_8b.value,
),
build_hf_repo_model_entry(
"llama-guard3:1b",
CoreModelId.llama_guard_3_1b.value,
),
ProviderModelEntry(
provider_model_id="all-minilm:l6-v2",
aliases=["all-minilm"],
@ -100,4 +103,4 @@ MODEL_ENTRIES = [
"context_length": 8192,
},
),
]
] + SAFETY_MODELS_ENTRIES

23
llama_stack/providers/remote/inference/openai/openai.py
View file

@ -8,7 +8,7 @@ import logging
from collections.abc import AsyncIterator
from typing import Any
from openai import AsyncOpenAI
from openai import AsyncOpenAI, NotFoundError
from llama_stack.apis.inference import (
OpenAIChatCompletion,
@ -60,6 +60,27 @@ class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
# litellm specific model names, an abstraction leak.
self.is_openai_compat = True
async def check_model_availability(self, model: str) -> bool:
"""
Check if a specific model is available from OpenAI.
:param model: The model identifier to check.
:return: True if the model is available dynamically, False otherwise.
"""
try:
openai_client = self._get_openai_client()
retrieved_model = await openai_client.models.retrieve(model)
logger.info(f"Model {retrieved_model.id} is available from OpenAI")
return True
except NotFoundError:
logger.error(f"Model {model} is not available from OpenAI")
return False
except Exception as e:
logger.error(f"Failed to check model availability from OpenAI: {e}")
return False
async def initialize(self) -> None:
await super().initialize()

1
llama_stack/providers/remote/vector_io/chroma/chroma.py
View file

@ -217,7 +217,6 @@ class ChromaVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
embedding_model: str | None = None,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorStoreObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Chroma")

8
llama_stack/providers/remote/vector_io/milvus/config.py
View file

@ -8,7 +8,7 @@ from typing import Any
from pydantic import BaseModel, ConfigDict, Field
from llama_stack.providers.utils.kvstore.config import KVStoreConfig
from llama_stack.providers.utils.kvstore.config import KVStoreConfig, SqliteKVStoreConfig
from llama_stack.schema_utils import json_schema_type
@ -17,7 +17,7 @@ class MilvusVectorIOConfig(BaseModel):
uri: str = Field(description="The URI of the Milvus server")
token: str | None = Field(description="The token of the Milvus server")
consistency_level: str = Field(description="The consistency level of the Milvus server", default="Strong")
kvstore: KVStoreConfig | None = Field(description="Config for KV store backend (SQLite only for now)", default=None)
kvstore: KVStoreConfig = Field(description="Config for KV store backend")
embedding_model: str | None = Field(
default=None,
description="Optional default embedding model for this provider. If not specified, will use system default.",
@ -36,6 +36,10 @@ class MilvusVectorIOConfig(BaseModel):
return {
"uri": "${env.MILVUS_ENDPOINT}",
"token": "${env.MILVUS_TOKEN}",
"kvstore": SqliteKVStoreConfig.sample_run_config(
__distro_dir__=__distro_dir__,
db_name="milvus_remote_registry.db",
),
# Optional: Configure default embedding model for this provider
# "embedding_model": "all-MiniLM-L6-v2",
# "embedding_dimension": 384, # Only needed for variable-dimension models

163
llama_stack/providers/remote/vector_io/milvus/milvus.py
View file

@ -12,7 +12,7 @@ import re
from typing import Any
from numpy.typing import NDArray
from pymilvus import DataType, MilvusClient
from pymilvus import DataType, Function, FunctionType, MilvusClient
from llama_stack.apis.files.files import Files
from llama_stack.apis.inference import Inference, InterleavedContent
@ -74,12 +74,66 @@ class MilvusIndex(EmbeddingIndex):
assert len(chunks) == len(embeddings), (
f"Chunk length {len(chunks)} does not match embedding length {len(embeddings)}"
)
if not await asyncio.to_thread(self.client.has_collection, self.collection_name):
logger.info(f"Creating new collection {self.collection_name} with nullable sparse field")
# Create schema for vector search
schema = self.client.create_schema()
schema.add_field(
field_name="chunk_id",
datatype=DataType.VARCHAR,
is_primary=True,
max_length=100,
)
schema.add_field(
field_name="content",
datatype=DataType.VARCHAR,
max_length=65535,
enable_analyzer=True, # Enable text analysis for BM25
)
schema.add_field(
field_name="vector",
datatype=DataType.FLOAT_VECTOR,
dim=len(embeddings[0]),
)
schema.add_field(
field_name="chunk_content",
datatype=DataType.JSON,
)
# Add sparse vector field for BM25 (required by the function)
schema.add_field(
field_name="sparse",
datatype=DataType.SPARSE_FLOAT_VECTOR,
)
# Create indexes
index_params = self.client.prepare_index_params()
index_params.add_index(
field_name="vector",
index_type="FLAT",
metric_type="COSINE",
)
# Add index for sparse field (required by BM25 function)
index_params.add_index(
field_name="sparse",
index_type="SPARSE_INVERTED_INDEX",
metric_type="BM25",
)
# Add BM25 function for full-text search
bm25_function = Function(
name="text_bm25_emb",
input_field_names=["content"],
output_field_names=["sparse"],
function_type=FunctionType.BM25,
)
schema.add_function(bm25_function)
await asyncio.to_thread(
self.client.create_collection,
self.collection_name,
dimension=len(embeddings[0]),
auto_id=True,
schema=schema,
index_params=index_params,
consistency_level=self.consistency_level,
)
@ -88,8 +142,10 @@ class MilvusIndex(EmbeddingIndex):
data.append(
{
"chunk_id": chunk.chunk_id,
"content": chunk.content,
"vector": embedding,
"chunk_content": chunk.model_dump(),
# sparse field will be handled by BM25 function automatically
}
)
try:
@ -107,6 +163,7 @@ class MilvusIndex(EmbeddingIndex):
self.client.search,
collection_name=self.collection_name,
data=[embedding],
anns_field="vector",
limit=k,
output_fields=["*"],
search_params={"params": {"radius": score_threshold}},
@ -121,7 +178,64 @@ class MilvusIndex(EmbeddingIndex):
k: int,
score_threshold: float,
) -> QueryChunksResponse:
raise NotImplementedError("Keyword search is not supported in Milvus")
"""
Perform BM25-based keyword search using Milvus's built-in full-text search.
"""
try:
# Use Milvus's built-in BM25 search
search_res = await asyncio.to_thread(
self.client.search,
collection_name=self.collection_name,
data=[query_string], # Raw text query
anns_field="sparse", # Use sparse field for BM25
output_fields=["chunk_content"], # Output the chunk content
limit=k,
search_params={
"params": {
"drop_ratio_search": 0.2, # Ignore low-importance terms
}
},
)
chunks = []
scores = []
for res in search_res[0]:
chunk = Chunk(**res["entity"]["chunk_content"])
chunks.append(chunk)
scores.append(res["distance"]) # BM25 score from Milvus
# Filter by score threshold
filtered_chunks = [chunk for chunk, score in zip(chunks, scores, strict=False) if score >= score_threshold]
filtered_scores = [score for score in scores if score >= score_threshold]
return QueryChunksResponse(chunks=filtered_chunks, scores=filtered_scores)
except Exception as e:
logger.error(f"Error performing BM25 search: {e}")
# Fallback to simple text search
return await self._fallback_keyword_search(query_string, k, score_threshold)
async def _fallback_keyword_search(
self,
query_string: str,
k: int,
score_threshold: float,
) -> QueryChunksResponse:
"""
Fallback to simple text search when BM25 search is not available.
"""
# Simple text search using content field
search_res = await asyncio.to_thread(
self.client.query,
collection_name=self.collection_name,
filter='content like "%{content}%"',
filter_params={"content": query_string},
output_fields=["*"],
limit=k,
)
chunks = [Chunk(**res["chunk_content"]) for res in search_res]
scores = [1.0] * len(chunks) # Simple binary score for text search
return QueryChunksResponse(chunks=chunks, scores=scores)
async def query_hybrid(
self,
@ -179,7 +293,8 @@ class MilvusVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolP
uri = os.path.expanduser(self.config.db_path)
self.client = MilvusClient(uri=uri)
self.openai_vector_stores = await self._load_openai_vector_stores()
# Load existing OpenAI vector stores into the in-memory cache
await self.initialize_openai_vector_stores()
async def shutdown(self) -> None:
self.client.close()
@ -246,38 +361,16 @@ class MilvusVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolP
if not index:
raise ValueError(f"Vector DB {vector_db_id} not found")
if params and params.get("mode") == "keyword":
# Check if this is inline Milvus (Milvus-Lite)
if hasattr(self.config, "db_path"):
raise NotImplementedError(
"Keyword search is not supported in Milvus-Lite. "
"Please use a remote Milvus server for keyword search functionality."
)
return await index.query_chunks(query, params)
async def _save_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Save vector store metadata to persistent storage."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
self.openai_vector_stores[store_id] = store_info
async def _update_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Update vector store metadata in persistent storage."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
self.openai_vector_stores[store_id] = store_info
async def _delete_openai_vector_store_from_storage(self, store_id: str) -> None:
"""Delete vector store metadata from persistent storage."""
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.delete(key)
if store_id in self.openai_vector_stores:
del self.openai_vector_stores[store_id]
async def _load_openai_vector_stores(self) -> dict[str, dict[str, Any]]:
"""Load all vector store metadata from persistent storage."""
assert self.kvstore is not None
start_key = OPENAI_VECTOR_STORES_PREFIX
end_key = f"{OPENAI_VECTOR_STORES_PREFIX}\xff"
stored = await self.kvstore.values_in_range(start_key, end_key)
return {json.loads(s)["id"]: json.loads(s) for s in stored}
async def _save_openai_vector_store_file(
self, store_id: str, file_id: str, file_info: dict[str, Any], file_contents: list[dict[str, Any]]
) -> None:

10
llama_stack/providers/remote/vector_io/pgvector/config.py
View file

@ -8,6 +8,10 @@ from typing import Any
from pydantic import BaseModel, Field
from llama_stack.providers.utils.kvstore.config import (
KVStoreConfig,
SqliteKVStoreConfig,
)
from llama_stack.schema_utils import json_schema_type
@ -18,6 +22,7 @@ class PGVectorVectorIOConfig(BaseModel):
db: str | None = Field(default="postgres")
user: str | None = Field(default="postgres")
password: str | None = Field(default="mysecretpassword")
kvstore: KVStoreConfig | None = Field(description="Config for KV store backend (SQLite only for now)", default=None)
embedding_model: str | None = Field(
default=None,
description="Optional default embedding model for this provider. If not specified, will use system default.",
@ -30,6 +35,7 @@ class PGVectorVectorIOConfig(BaseModel):
@classmethod
def sample_run_config(
cls,
__distro_dir__: str,
host: str = "${env.PGVECTOR_HOST:=localhost}",
port: int | str = "${env.PGVECTOR_PORT:=5432}",
db: str = "${env.PGVECTOR_DB}",
@ -43,6 +49,10 @@ class PGVectorVectorIOConfig(BaseModel):
"db": db,
"user": user,
"password": password,
"kvstore": SqliteKVStoreConfig.sample_run_config(
__distro_dir__=__distro_dir__,
db_name="pgvector_registry.db",
),
# Optional: Configure default embedding model for this provider
# "embedding_model": "all-MiniLM-L6-v2",
# "embedding_dimension": 384, # Only needed for variable-dimension models

280
llama_stack/providers/remote/vector_io/pgvector/pgvector.py
View file

@ -13,24 +13,18 @@ from psycopg2 import sql
from psycopg2.extras import Json, execute_values
from pydantic import BaseModel, TypeAdapter
from llama_stack.apis.files.files import Files
from llama_stack.apis.inference import InterleavedContent
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.vector_io import (
Chunk,
QueryChunksResponse,
SearchRankingOptions,
VectorIO,
VectorStoreChunkingStrategy,
VectorStoreDeleteResponse,
VectorStoreFileContentsResponse,
VectorStoreFileObject,
VectorStoreFileStatus,
VectorStoreListFilesResponse,
VectorStoreListResponse,
VectorStoreObject,
VectorStoreSearchResponsePage,
)
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.utils.kvstore import kvstore_impl
from llama_stack.providers.utils.kvstore.api import KVStore
from llama_stack.providers.utils.memory.openai_vector_store_mixin import OpenAIVectorStoreMixin
from llama_stack.providers.utils.memory.vector_store import (
EmbeddingIndex,
VectorDBWithIndex,
@ -40,6 +34,13 @@ from .config import PGVectorVectorIOConfig
log = logging.getLogger(__name__)
VERSION = "v3"
VECTOR_DBS_PREFIX = f"vector_dbs:pgvector:{VERSION}::"
VECTOR_INDEX_PREFIX = f"vector_index:pgvector:{VERSION}::"
OPENAI_VECTOR_STORES_PREFIX = f"openai_vector_stores:pgvector:{VERSION}::"
OPENAI_VECTOR_STORES_FILES_PREFIX = f"openai_vector_stores_files:pgvector:{VERSION}::"
OPENAI_VECTOR_STORES_FILES_CONTENTS_PREFIX = f"openai_vector_stores_files_contents:pgvector:{VERSION}::"
def check_extension_version(cur):
cur.execute("SELECT extversion FROM pg_extension WHERE extname = 'vector'")
@ -69,7 +70,7 @@ def load_models(cur, cls):
class PGVectorIndex(EmbeddingIndex):
def __init__(self, vector_db: VectorDB, dimension: int, conn):
def __init__(self, vector_db: VectorDB, dimension: int, conn, kvstore: KVStore | None = None):
self.conn = conn
with conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
# Sanitize the table name by replacing hyphens with underscores
@ -77,6 +78,7 @@ class PGVectorIndex(EmbeddingIndex):
# when created with patterns like "test-vector-db-{uuid4()}"
sanitized_identifier = vector_db.identifier.replace("-", "_")
self.table_name = f"vector_store_{sanitized_identifier}"
self.kvstore = kvstore
cur.execute(
f"""
@ -158,15 +160,28 @@ class PGVectorIndex(EmbeddingIndex):
cur.execute(f"DROP TABLE IF EXISTS {self.table_name}")
class PGVectorVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
def __init__(self, config: PGVectorVectorIOConfig, inference_api: Api.inference) -> None:
class PGVectorVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPrivate):
def __init__(
self,
config: PGVectorVectorIOConfig,
inference_api: Api.inference,
files_api: Files | None = None,
) -> None:
self.config = config
self.inference_api = inference_api
self.conn = None
self.cache = {}
self.files_api = files_api
self.kvstore: KVStore | None = None
self.vector_db_store = None
self.openai_vector_store: dict[str, dict[str, Any]] = {}
self.metadatadata_collection_name = "openai_vector_stores_metadata"
async def initialize(self) -> None:
log.info(f"Initializing PGVector memory adapter with config: {self.config}")
self.kvstore = await kvstore_impl(self.config.kvstore)
await self.initialize_openai_vector_stores()
try:
self.conn = psycopg2.connect(
host=self.config.host,
@ -201,14 +216,28 @@ class PGVectorVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
log.info("Connection to PGVector database server closed")
async def register_vector_db(self, vector_db: VectorDB) -> None:
# Persist vector DB metadata in the KV store
assert self.kvstore is not None
# Upsert model metadata in Postgres
upsert_models(self.conn, [(vector_db.identifier, vector_db)])
index = PGVectorIndex(vector_db, vector_db.embedding_dimension, self.conn)
self.cache[vector_db.identifier] = VectorDBWithIndex(vector_db, index, self.inference_api)
# Create and cache the PGVector index table for the vector DB
index = VectorDBWithIndex(
vector_db,
index=PGVectorIndex(vector_db, vector_db.embedding_dimension, self.conn, kvstore=self.kvstore),
inference_api=self.inference_api,
)
self.cache[vector_db.identifier] = index
async def unregister_vector_db(self, vector_db_id: str) -> None:
await self.cache[vector_db_id].index.delete()
del self.cache[vector_db_id]
# Remove provider index and cache
if vector_db_id in self.cache:
await self.cache[vector_db_id].index.delete()
del self.cache[vector_db_id]
# Delete vector DB metadata from KV store
assert self.kvstore is not None
await self.kvstore.delete(key=f"{VECTOR_DBS_PREFIX}{vector_db_id}")
async def insert_chunks(
self,
@ -237,107 +266,124 @@ class PGVectorVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
self.cache[vector_db_id] = VectorDBWithIndex(vector_db, index, self.inference_api)
return self.cache[vector_db_id]
async def openai_create_vector_store(
self,
name: str,
file_ids: list[str] | None = None,
expires_after: dict[str, Any] | None = None,
chunking_strategy: dict[str, Any] | None = None,
metadata: dict[str, Any] | None = None,
embedding_model: str | None = None,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorStoreObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
# OpenAI Vector Stores File operations are not supported in PGVector
async def _save_openai_vector_store_file(
self, store_id: str, file_id: str, file_info: dict[str, Any], file_contents: list[dict[str, Any]]
) -> None:
"""Save vector store file metadata to Postgres database."""
if self.conn is None:
raise RuntimeError("PostgreSQL connection is not initialized")
try:
with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
cur.execute(
"""
CREATE TABLE IF NOT EXISTS openai_vector_store_files (
store_id TEXT,
file_id TEXT,
metadata JSONB,
PRIMARY KEY (store_id, file_id)
)
"""
)
cur.execute(
"""
CREATE TABLE IF NOT EXISTS openai_vector_store_files_contents (
store_id TEXT,
file_id TEXT,
contents JSONB,
PRIMARY KEY (store_id, file_id)
)
"""
)
# Insert file metadata
files_query = sql.SQL(
"""
INSERT INTO openai_vector_store_files (store_id, file_id, metadata)
VALUES %s
ON CONFLICT (store_id, file_id) DO UPDATE SET metadata = EXCLUDED.metadata
"""
)
files_values = [(store_id, file_id, Json(file_info))]
execute_values(cur, files_query, files_values, template="(%s, %s, %s)")
# Insert file contents
contents_query = sql.SQL(
"""
INSERT INTO openai_vector_store_files_contents (store_id, file_id, contents)
VALUES %s
ON CONFLICT (store_id, file_id) DO UPDATE SET contents = EXCLUDED.contents
"""
)
contents_values = [(store_id, file_id, Json(file_contents))]
execute_values(cur, contents_query, contents_values, template="(%s, %s, %s)")
except Exception as e:
log.error(f"Error saving openai vector store file {file_id} for store {store_id}: {e}")
raise
async def openai_list_vector_stores(
self,
limit: int | None = 20,
order: str | None = "desc",
after: str | None = None,
before: str | None = None,
) -> VectorStoreListResponse:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def _load_openai_vector_store_file(self, store_id: str, file_id: str) -> dict[str, Any]:
"""Load vector store file metadata from Postgres database."""
if self.conn is None:
raise RuntimeError("PostgreSQL connection is not initialized")
try:
with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
cur.execute(
"SELECT metadata FROM openai_vector_store_files WHERE store_id = %s AND file_id = %s",
(store_id, file_id),
)
row = cur.fetchone()
return row[0] if row and row[0] is not None else {}
except Exception as e:
log.error(f"Error loading openai vector store file {file_id} for store {store_id}: {e}")
return {}
async def openai_retrieve_vector_store(
self,
vector_store_id: str,
) -> VectorStoreObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def _load_openai_vector_store_file_contents(self, store_id: str, file_id: str) -> list[dict[str, Any]]:
"""Load vector store file contents from Postgres database."""
if self.conn is None:
raise RuntimeError("PostgreSQL connection is not initialized")
try:
with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
cur.execute(
"SELECT contents FROM openai_vector_store_files_contents WHERE store_id = %s AND file_id = %s",
(store_id, file_id),
)
row = cur.fetchone()
return row[0] if row and row[0] is not None else []
except Exception as e:
log.error(f"Error loading openai vector store file contents for {file_id} in store {store_id}: {e}")
return []
async def openai_update_vector_store(
self,
vector_store_id: str,
name: str | None = None,
expires_after: dict[str, Any] | None = None,
metadata: dict[str, Any] | None = None,
) -> VectorStoreObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def _update_openai_vector_store_file(self, store_id: str, file_id: str, file_info: dict[str, Any]) -> None:
"""Update vector store file metadata in Postgres database."""
if self.conn is None:
raise RuntimeError("PostgreSQL connection is not initialized")
try:
with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
query = sql.SQL(
"""
INSERT INTO openai_vector_store_files (store_id, file_id, metadata)
VALUES %s
ON CONFLICT (store_id, file_id) DO UPDATE SET metadata = EXCLUDED.metadata
"""
)
values = [(store_id, file_id, Json(file_info))]
execute_values(cur, query, values, template="(%s, %s, %s)")
except Exception as e:
log.error(f"Error updating openai vector store file {file_id} for store {store_id}: {e}")
raise
async def openai_delete_vector_store(
self,
vector_store_id: str,
) -> VectorStoreDeleteResponse:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def openai_search_vector_store(
self,
vector_store_id: str,
query: str | list[str],
filters: dict[str, Any] | None = None,
max_num_results: int | None = 10,
ranking_options: SearchRankingOptions | None = None,
rewrite_query: bool | None = False,
search_mode: str | None = "vector",
) -> VectorStoreSearchResponsePage:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
chunking_strategy: VectorStoreChunkingStrategy | None = None,
) -> VectorStoreFileObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def openai_list_files_in_vector_store(
self,
vector_store_id: str,
limit: int | None = 20,
order: str | None = "desc",
after: str | None = None,
before: str | None = None,
filter: VectorStoreFileStatus | None = None,
) -> VectorStoreListFilesResponse:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def openai_retrieve_vector_store_file(
self,
vector_store_id: str,
file_id: str,
) -> VectorStoreFileObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def openai_retrieve_vector_store_file_contents(
self,
vector_store_id: str,
file_id: str,
) -> VectorStoreFileContentsResponse:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def openai_update_vector_store_file(
self,
vector_store_id: str,
file_id: str,
attributes: dict[str, Any] | None = None,
) -> VectorStoreFileObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def openai_delete_vector_store_file(
self,
vector_store_id: str,
file_id: str,
) -> VectorStoreFileObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in PGVector")
async def _delete_openai_vector_store_file_from_storage(self, store_id: str, file_id: str) -> None:
"""Delete vector store file metadata from Postgres database."""
if self.conn is None:
raise RuntimeError("PostgreSQL connection is not initialized")
try:
with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
cur.execute(
"DELETE FROM openai_vector_store_files WHERE store_id = %s AND file_id = %s",
(store_id, file_id),
)
cur.execute(
"DELETE FROM openai_vector_store_files_contents WHERE store_id = %s AND file_id = %s",
(store_id, file_id),
)
except Exception as e:
log.error(f"Error deleting openai vector store file {file_id} for store {store_id}: {e}")
raise

1
llama_stack/providers/remote/vector_io/qdrant/qdrant.py
View file

@ -214,7 +214,6 @@ class QdrantVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
embedding_model: str | None = None,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorStoreObject:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Qdrant")

12
llama_stack/providers/remote/vector_io/weaviate/config.py
View file

@ -8,10 +8,16 @@ from typing import Any
from pydantic import BaseModel, Field
from llama_stack.providers.utils.kvstore.config import (
KVStoreConfig,
SqliteKVStoreConfig,
)
class WeaviateRequestProviderData(BaseModel):
weaviate_api_key: str
weaviate_cluster_url: str
kvstore: KVStoreConfig | None = Field(description="Config for KV store backend (SQLite only for now)", default=None)
class WeaviateVectorIOConfig(BaseModel):
@ -25,8 +31,12 @@ class WeaviateVectorIOConfig(BaseModel):
)
@classmethod
def sample_run_config(cls, **kwargs: Any) -> dict[str, Any]:
def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> dict[str, Any]:
return {
"kvstore": SqliteKVStoreConfig.sample_run_config(
__distro_dir__=__distro_dir__,
db_name="weaviate_registry.db",
),
# Optional: Configure default embedding model for this provider
# "embedding_model": "all-MiniLM-L6-v2",
# "embedding_dimension": 384, # Only needed for variable-dimension models

64
llama_stack/providers/remote/vector_io/weaviate/weaviate.py
View file

@ -14,10 +14,13 @@ from weaviate.classes.init import Auth
from weaviate.classes.query import Filter
from llama_stack.apis.common.content_types import InterleavedContent
from llama_stack.apis.files.files import Files
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.vector_io import Chunk, QueryChunksResponse, VectorIO
from llama_stack.distribution.request_headers import NeedsRequestProviderData
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.utils.kvstore import kvstore_impl
from llama_stack.providers.utils.kvstore.api import KVStore
from llama_stack.providers.utils.memory.vector_store import (
EmbeddingIndex,
VectorDBWithIndex,
@ -27,11 +30,19 @@ from .config import WeaviateRequestProviderData, WeaviateVectorIOConfig
log = logging.getLogger(__name__)
VERSION = "v3"
VECTOR_DBS_PREFIX = f"vector_dbs:weaviate:{VERSION}::"
VECTOR_INDEX_PREFIX = f"vector_index:weaviate:{VERSION}::"
OPENAI_VECTOR_STORES_PREFIX = f"openai_vector_stores:weaviate:{VERSION}::"
OPENAI_VECTOR_STORES_FILES_PREFIX = f"openai_vector_stores_files:weaviate:{VERSION}::"
OPENAI_VECTOR_STORES_FILES_CONTENTS_PREFIX = f"openai_vector_stores_files_contents:weaviate:{VERSION}::"
class WeaviateIndex(EmbeddingIndex):
def __init__(self, client: weaviate.Client, collection_name: str):
def __init__(self, client: weaviate.Client, collection_name: str, kvstore: KVStore | None = None):
self.client = client
self.collection_name = collection_name
self.kvstore = kvstore
async def add_chunks(self, chunks: list[Chunk], embeddings: NDArray):
assert len(chunks) == len(embeddings), (
@ -109,11 +120,21 @@ class WeaviateVectorIOAdapter(
NeedsRequestProviderData,
VectorDBsProtocolPrivate,
):
def __init__(self, config: WeaviateVectorIOConfig, inference_api: Api.inference) -> None:
def __init__(
self,
config: WeaviateVectorIOConfig,
inference_api: Api.inference,
files_api: Files | None,
) -> None:
self.config = config
self.inference_api = inference_api
self.client_cache = {}
self.cache = {}
self.files_api = files_api
self.kvstore: KVStore | None = None
self.vector_db_store = None
self.openai_vector_stores: dict[str, dict[str, Any]] = {}
self.metadata_collection_name = "openai_vector_stores_metadata"
def _get_client(self) -> weaviate.Client:
provider_data = self.get_request_provider_data()
@ -132,7 +153,26 @@ class WeaviateVectorIOAdapter(
return client
async def initialize(self) -> None:
pass
"""Set up KV store and load existing vector DBs and OpenAI vector stores."""
# Initialize KV store for metadata
self.kvstore = await kvstore_impl(self.config.kvstore)
# Load existing vector DB definitions
start_key = VECTOR_DBS_PREFIX
end_key = f"{VECTOR_DBS_PREFIX}\xff"
stored = await self.kvstore.values_in_range(start_key, end_key)
for raw in stored:
vector_db = VectorDB.model_validate_json(raw)
client = self._get_client()
idx = WeaviateIndex(client=client, collection_name=vector_db.identifier, kvstore=self.kvstore)
self.cache[vector_db.identifier] = VectorDBWithIndex(
vector_db=vector_db,
index=idx,
inference_api=self.inference_api,
)
# Load OpenAI vector stores metadata into cache
await self.initialize_openai_vector_stores()
async def shutdown(self) -> None:
for client in self.client_cache.values():
@ -206,3 +246,21 @@ class WeaviateVectorIOAdapter(
raise ValueError(f"Vector DB {vector_db_id} not found")
return await index.query_chunks(query, params)
# OpenAI Vector Stores File operations are not supported in Weaviate
async def _save_openai_vector_store_file(
self, store_id: str, file_id: str, file_info: dict[str, Any], file_contents: list[dict[str, Any]]
) -> None:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Weaviate")
async def _load_openai_vector_store_file(self, store_id: str, file_id: str) -> dict[str, Any]:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Weaviate")
async def _load_openai_vector_store_file_contents(self, store_id: str, file_id: str) -> list[dict[str, Any]]:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Weaviate")
async def _update_openai_vector_store_file(self, store_id: str, file_id: str, file_info: dict[str, Any]) -> None:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Weaviate")
async def _delete_openai_vector_store_file_from_storage(self, store_id: str, file_id: str) -> None:
raise NotImplementedError("OpenAI Vector Stores API is not supported in Weaviate")

8
llama_stack/providers/utils/inference/litellm_openai_mixin.py
View file

@ -13,7 +13,6 @@ from llama_stack.apis.common.content_types import (
InterleavedContent,
InterleavedContentItem,
)
from llama_stack.apis.common.errors import UnsupportedModelError
from llama_stack.apis.inference import (
ChatCompletionRequest,
ChatCompletionResponse,
@ -39,7 +38,6 @@ from llama_stack.apis.inference import (
ToolDefinition,
ToolPromptFormat,
)
from llama_stack.apis.models import Model
from llama_stack.distribution.request_headers import NeedsRequestProviderData
from llama_stack.log import get_logger
from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper
@ -90,12 +88,6 @@ class LiteLLMOpenAIMixin(
async def shutdown(self):
pass
async def register_model(self, model: Model) -> Model:
model_id = self.get_provider_model_id(model.provider_resource_id)
if model_id is None:
raise UnsupportedModelError(model.provider_resource_id, self.alias_to_provider_id_map.keys())
return model
def get_litellm_model_name(self, model_id: str) -> str:
# users may be using openai/ prefix in their model names. the openai/models.py did this by default.
# model_id.startswith("openai/") is for backwards compatibility.

68
llama_stack/providers/utils/memory/openai_vector_store_mixin.py
View file

@ -5,6 +5,7 @@
# the root directory of this source tree.
import asyncio
import json
import logging
import mimetypes
import time
@ -35,6 +36,7 @@ from llama_stack.apis.vector_io import (
VectorStoreSearchResponse,
VectorStoreSearchResponsePage,
)
from llama_stack.providers.utils.kvstore.api import KVStore
from llama_stack.providers.utils.memory.vector_store import content_from_data_and_mime_type, make_overlapped_chunks
logger = logging.getLogger(__name__)
@ -59,26 +61,45 @@ class OpenAIVectorStoreMixin(ABC):
# These should be provided by the implementing class
openai_vector_stores: dict[str, dict[str, Any]]
files_api: Files | None
# KV store for persisting OpenAI vector store metadata
kvstore: KVStore | None
@abstractmethod
async def _save_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Save vector store metadata to persistent storage."""
pass
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
# update in-memory cache
self.openai_vector_stores[store_id] = store_info
@abstractmethod
async def _load_openai_vector_stores(self) -> dict[str, dict[str, Any]]:
"""Load all vector store metadata from persistent storage."""
pass
assert self.kvstore is not None
start_key = OPENAI_VECTOR_STORES_PREFIX
end_key = f"{OPENAI_VECTOR_STORES_PREFIX}\xff"
stored_data = await self.kvstore.values_in_range(start_key, end_key)
stores: dict[str, dict[str, Any]] = {}
for item in stored_data:
info = json.loads(item)
stores[info["id"]] = info
return stores
@abstractmethod
async def _update_openai_vector_store(self, store_id: str, store_info: dict[str, Any]) -> None:
"""Update vector store metadata in persistent storage."""
pass
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.set(key=key, value=json.dumps(store_info))
# update in-memory cache
self.openai_vector_stores[store_id] = store_info
@abstractmethod
async def _delete_openai_vector_store_from_storage(self, store_id: str) -> None:
"""Delete vector store metadata from persistent storage."""
pass
assert self.kvstore is not None
key = f"{OPENAI_VECTOR_STORES_PREFIX}{store_id}"
await self.kvstore.delete(key)
# remove from in-memory cache
self.openai_vector_stores.pop(store_id, None)
@abstractmethod
async def _save_openai_vector_store_file(
@ -117,6 +138,10 @@ class OpenAIVectorStoreMixin(ABC):
"""Unregister a vector database (provider-specific implementation)."""
pass
async def initialize_openai_vector_stores(self) -> None:
"""Load existing OpenAI vector stores into the in-memory cache."""
self.openai_vector_stores = await self._load_openai_vector_stores()
@abstractmethod
async def insert_chunks(
self,
@ -147,8 +172,9 @@ class OpenAIVectorStoreMixin(ABC):
provider_vector_db_id: str | None = None,
) -> VectorStoreObject:
"""Creates a vector store."""
store_id = name or str(uuid.uuid4())
created_at = int(time.time())
# Derive the canonical vector_db_id (allow override, else generate)
vector_db_id = provider_vector_db_id or f"vs_{uuid.uuid4()}"
if provider_id is None:
raise ValueError("Provider ID is required")
@ -156,19 +182,19 @@ class OpenAIVectorStoreMixin(ABC):
if embedding_model is None:
raise ValueError("Embedding model is required")
# Use provided embedding dimension or default to 384
# Embedding dimension is required (defaulted to 384 if not provided)
if embedding_dimension is None:
raise ValueError("Embedding dimension is required")
provider_vector_db_id = provider_vector_db_id or store_id
# Register the VectorDB backing this vector store
vector_db = VectorDB(
identifier=store_id,
identifier=vector_db_id,
embedding_dimension=embedding_dimension,
embedding_model=embedding_model,
provider_id=provider_id,
provider_resource_id=provider_vector_db_id,
provider_resource_id=vector_db_id,
vector_db_name=name,
)
# Register the vector DB
await self.register_vector_db(vector_db)
# Create OpenAI vector store metadata
@ -182,11 +208,11 @@ class OpenAIVectorStoreMixin(ABC):
in_progress=0,
total=0,
)
store_info = {
"id": store_id,
store_info: dict[str, Any] = {
"id": vector_db_id,
"object": "vector_store",
"created_at": created_at,
"name": store_id,
"name": name,
"usage_bytes": 0,
"file_counts": file_counts.model_dump(),
"status": status,
@ -206,18 +232,18 @@ class OpenAIVectorStoreMixin(ABC):
store_info["metadata"] = metadata
# Save to persistent storage (provider-specific)
await self._save_openai_vector_store(store_id, store_info)
await self._save_openai_vector_store(vector_db_id, store_info)
# Store in memory cache
self.openai_vector_stores[store_id] = store_info
self.openai_vector_stores[vector_db_id] = store_info
# Now that our vector store is created, attach any files that were provided
file_ids = file_ids or []
tasks = [self.openai_attach_file_to_vector_store(store_id, file_id) for file_id in file_ids]
tasks = [self.openai_attach_file_to_vector_store(vector_db_id, file_id) for file_id in file_ids]
await asyncio.gather(*tasks)
# Get the updated store info and return it
store_info = self.openai_vector_stores[store_id]
store_info = self.openai_vector_stores[vector_db_id]
return VectorStoreObject.model_validate(store_info)
async def openai_list_vector_stores(

2
llama_stack/templates/nvidia/nvidia.py
View file

@ -68,7 +68,7 @@ def get_distribution_template() -> DistributionTemplate:
),
]
default_models = get_model_registry(available_models)
default_models, _ = get_model_registry(available_models)
return DistributionTemplate(
name="nvidia",
distro_type="self_hosted",

4
llama_stack/templates/open-benchmark/open_benchmark.py
View file

@ -128,6 +128,7 @@ def get_distribution_template() -> DistributionTemplate:
provider_id="${env.ENABLE_PGVECTOR:+pgvector}",
provider_type="remote::pgvector",
config=PGVectorVectorIOConfig.sample_run_config(
f"~/.llama/distributions/{name}",
db="${env.PGVECTOR_DB:=}",
user="${env.PGVECTOR_USER:=}",
password="${env.PGVECTOR_PASSWORD:=}",
@ -146,7 +147,8 @@ def get_distribution_template() -> DistributionTemplate:
),
]
default_models = get_model_registry(available_models) + [
models, _ = get_model_registry(available_models)
default_models = models + [
ModelInput(
model_id="meta-llama/Llama-3.3-70B-Instruct",
provider_id="groq",

3
llama_stack/templates/open-benchmark/run.yaml
View file

@ -54,6 +54,9 @@ providers:
db: ${env.PGVECTOR_DB:=}
user: ${env.PGVECTOR_USER:=}
password: ${env.PGVECTOR_PASSWORD:=}
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/open-benchmark}/pgvector_registry.db
safety:
- provider_id: llama-guard
provider_type: inline::llama-guard

23
llama_stack/templates/starter/run.yaml
View file

@ -166,6 +166,9 @@ providers:
db: ${env.PGVECTOR_DB:=}
user: ${env.PGVECTOR_USER:=}
password: ${env.PGVECTOR_PASSWORD:=}
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/pgvector_registry.db
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
@ -1171,24 +1174,8 @@ models:
provider_id: ${env.ENABLE_SENTENCE_TRANSFORMERS:=sentence-transformers}
model_type: embedding
shields:
- shield_id: ${env.ENABLE_OLLAMA:=__disabled__}
provider_id: llama-guard
provider_shield_id: ${env.ENABLE_OLLAMA:=__disabled__}/${env.SAFETY_MODEL:=llama-guard3:1b}
- shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}
provider_id: llama-guard
provider_shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}/${env.SAFETY_MODEL:=accounts/fireworks/models/llama-guard-3-8b}
- shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}
provider_id: llama-guard
provider_shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}/${env.SAFETY_MODEL:=accounts/fireworks/models/llama-guard-3-11b-vision}
- shield_id: ${env.ENABLE_TOGETHER:=__disabled__}
provider_id: llama-guard
provider_shield_id: ${env.ENABLE_TOGETHER:=__disabled__}/${env.SAFETY_MODEL:=meta-llama/Llama-Guard-3-8B}
- shield_id: ${env.ENABLE_TOGETHER:=__disabled__}
provider_id: llama-guard
provider_shield_id: ${env.ENABLE_TOGETHER:=__disabled__}/${env.SAFETY_MODEL:=meta-llama/Llama-Guard-3-11B-Vision-Turbo}
- shield_id: ${env.ENABLE_SAMBANOVA:=__disabled__}
provider_id: llama-guard
provider_shield_id: ${env.ENABLE_SAMBANOVA:=__disabled__}/${env.SAFETY_MODEL:=sambanova/Meta-Llama-Guard-3-8B}
- shield_id: ${env.SAFETY_MODEL:=__disabled__}
provider_shield_id: ${env.ENABLE_OLLAMA:=__disabled__}/${env.SAFETY_MODEL:=__disabled__}
vector_dbs: []
datasets: []
scoring_fns: []

95
llama_stack/templates/starter/starter.py
View file

@ -12,7 +12,6 @@ from llama_stack.distribution.datatypes import (
ModelInput,
Provider,
ProviderSpec,
ShieldInput,
ToolGroupInput,
)
from llama_stack.distribution.utils.dynamic import instantiate_class_type
@ -32,75 +31,39 @@ from llama_stack.providers.registry.inference import available_providers
from llama_stack.providers.remote.inference.anthropic.models import (
MODEL_ENTRIES as ANTHROPIC_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.anthropic.models import (
SAFETY_MODELS_ENTRIES as ANTHROPIC_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.bedrock.models import (
MODEL_ENTRIES as BEDROCK_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.bedrock.models import (
SAFETY_MODELS_ENTRIES as BEDROCK_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.cerebras.models import (
MODEL_ENTRIES as CEREBRAS_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.cerebras.models import (
SAFETY_MODELS_ENTRIES as CEREBRAS_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.databricks.databricks import (
MODEL_ENTRIES as DATABRICKS_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.databricks.databricks import (
SAFETY_MODELS_ENTRIES as DATABRICKS_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.fireworks.models import (
MODEL_ENTRIES as FIREWORKS_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.fireworks.models import (
SAFETY_MODELS_ENTRIES as FIREWORKS_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.gemini.models import (
MODEL_ENTRIES as GEMINI_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.gemini.models import (
SAFETY_MODELS_ENTRIES as GEMINI_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.groq.models import (
MODEL_ENTRIES as GROQ_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.groq.models import (
SAFETY_MODELS_ENTRIES as GROQ_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.nvidia.models import (
MODEL_ENTRIES as NVIDIA_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.nvidia.models import (
SAFETY_MODELS_ENTRIES as NVIDIA_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.openai.models import (
MODEL_ENTRIES as OPENAI_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.openai.models import (
SAFETY_MODELS_ENTRIES as OPENAI_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.runpod.runpod import (
MODEL_ENTRIES as RUNPOD_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.runpod.runpod import (
SAFETY_MODELS_ENTRIES as RUNPOD_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.sambanova.models import (
MODEL_ENTRIES as SAMBANOVA_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.sambanova.models import (
SAFETY_MODELS_ENTRIES as SAMBANOVA_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.inference.together.models import (
MODEL_ENTRIES as TOGETHER_MODEL_ENTRIES,
)
from llama_stack.providers.remote.inference.together.models import (
SAFETY_MODELS_ENTRIES as TOGETHER_SAFETY_MODELS_ENTRIES,
)
from llama_stack.providers.remote.vector_io.chroma.config import ChromaVectorIOConfig
from llama_stack.providers.remote.vector_io.pgvector.config import (
PGVectorVectorIOConfig,
@ -111,6 +74,7 @@ from llama_stack.templates.template import (
DistributionTemplate,
RunConfigSettings,
get_model_registry,
get_shield_registry,
)
@ -164,28 +128,13 @@ def _get_model_entries_for_provider(provider_type: str) -> list[ProviderModelEnt
def _get_model_safety_entries_for_provider(provider_type: str) -> list[ProviderModelEntry]:
"""Get model entries for a specific provider type."""
safety_model_entries_map = {
"openai": OPENAI_SAFETY_MODELS_ENTRIES,
"fireworks": FIREWORKS_SAFETY_MODELS_ENTRIES,
"together": TOGETHER_SAFETY_MODELS_ENTRIES,
"anthropic": ANTHROPIC_SAFETY_MODELS_ENTRIES,
"gemini": GEMINI_SAFETY_MODELS_ENTRIES,
"groq": GROQ_SAFETY_MODELS_ENTRIES,
"sambanova": SAMBANOVA_SAFETY_MODELS_ENTRIES,
"cerebras": CEREBRAS_SAFETY_MODELS_ENTRIES,
"bedrock": BEDROCK_SAFETY_MODELS_ENTRIES,
"databricks": DATABRICKS_SAFETY_MODELS_ENTRIES,
"nvidia": NVIDIA_SAFETY_MODELS_ENTRIES,
"runpod": RUNPOD_SAFETY_MODELS_ENTRIES,
}
# Special handling for providers with dynamic model entries
if provider_type == "ollama":
return [
"ollama": [
ProviderModelEntry(
provider_model_id="llama-guard3:1b",
provider_model_id="${env.SAFETY_MODEL:=__disabled__}",
model_type=ModelType.llm,
),
]
],
}
return safety_model_entries_map.get(provider_type, [])
@ -246,28 +195,20 @@ def get_remote_inference_providers() -> tuple[list[Provider], dict[str, list[Pro
# build a list of shields for all possible providers
def get_shields_for_providers(providers: list[Provider]) -> list[ShieldInput]:
shields = []
def get_safety_models_for_providers(providers: list[Provider]) -> dict[str, list[ProviderModelEntry]]:
available_models = {}
for provider in providers:
provider_type = provider.provider_type.split("::")[1]
safety_model_entries = _get_model_safety_entries_for_provider(provider_type)
if len(safety_model_entries) == 0:
continue
if provider.provider_id:
shield_id = provider.provider_id
else:
raise ValueError(f"Provider {provider.provider_type} has no provider_id")
for safety_model_entry in safety_model_entries:
print(f"provider.provider_id: {provider.provider_id}")
print(f"safety_model_entry.provider_model_id: {safety_model_entry.provider_model_id}")
shields.append(
ShieldInput(
provider_id="llama-guard",
shield_id=shield_id,
provider_shield_id=f"{provider.provider_id}/${{env.SAFETY_MODEL:={safety_model_entry.provider_model_id}}}",
)
)
return shields
env_var = f"ENABLE_{provider_type.upper().replace('-', '_').replace('::', '_')}"
provider_id = f"${{env.{env_var}:=__disabled__}}"
available_models[provider_id] = safety_model_entries
return available_models
def get_distribution_template() -> DistributionTemplate:
@ -300,6 +241,7 @@ def get_distribution_template() -> DistributionTemplate:
provider_id="${env.ENABLE_PGVECTOR:=__disabled__}",
provider_type="remote::pgvector",
config=PGVectorVectorIOConfig.sample_run_config(
f"~/.llama/distributions/{name}",
db="${env.PGVECTOR_DB:=}",
user="${env.PGVECTOR_USER:=}",
password="${env.PGVECTOR_PASSWORD:=}",
@ -307,8 +249,6 @@ def get_distribution_template() -> DistributionTemplate:
),
]
shields = get_shields_for_providers(remote_inference_providers)
providers = {
"inference": ([p.provider_type for p in remote_inference_providers] + ["inline::sentence-transformers"]),
"vector_io": ([p.provider_type for p in vector_io_providers]),
@ -361,7 +301,10 @@ def get_distribution_template() -> DistributionTemplate:
},
)
default_models = get_model_registry(available_models)
default_models, ids_conflict_in_models = get_model_registry(available_models)
available_safety_models = get_safety_models_for_providers(remote_inference_providers)
shields = get_shield_registry(available_safety_models, ids_conflict_in_models)
return DistributionTemplate(
name=name,

47
llama_stack/templates/template.py
View file

@ -37,7 +37,7 @@ from llama_stack.providers.utils.sqlstore.sqlstore import get_pip_packages as ge
def get_model_registry(
available_models: dict[str, list[ProviderModelEntry]],
) -> list[ModelInput]:
) -> tuple[list[ModelInput], bool]:
models = []
# check for conflicts in model ids
@ -74,7 +74,50 @@ def get_model_registry(
metadata=entry.metadata,
)
)
return models
return models, ids_conflict
def get_shield_registry(
available_safety_models: dict[str, list[ProviderModelEntry]],
ids_conflict_in_models: bool,
) -> list[ShieldInput]:
shields = []
# check for conflicts in shield ids
all_ids = set()
ids_conflict = False
for _, entries in available_safety_models.items():
for entry in entries:
ids = [entry.provider_model_id] + entry.aliases
for model_id in ids:
if model_id in all_ids:
ids_conflict = True
rich.print(
f"[yellow]Shield id {model_id} conflicts; all shield ids will be prefixed with provider id[/yellow]"
)
break
all_ids.update(ids)
if ids_conflict:
break
if ids_conflict:
break
for provider_id, entries in available_safety_models.items():
for entry in entries:
ids = [entry.provider_model_id] + entry.aliases
for model_id in ids:
identifier = f"{provider_id}/{model_id}" if ids_conflict and provider_id not in model_id else model_id
shields.append(
ShieldInput(
shield_id=identifier,
provider_shield_id=f"{provider_id}/{entry.provider_model_id}"
if ids_conflict_in_models
else entry.provider_model_id,
)
)
return shields
class DefaultModel(BaseModel):

2
llama_stack/templates/watsonx/watsonx.py
View file

@ -69,7 +69,7 @@ def get_distribution_template() -> DistributionTemplate:
},
)
default_models = get_model_registry(available_models)
default_models, _ = get_model_registry(available_models)
return DistributionTemplate(
name="watsonx",
distro_type="remote_hosted",

2
llama_stack/ui/package.json
View file

@ -20,7 +20,7 @@
"@radix-ui/react-tooltip": "^1.2.6",
"class-variance-authority": "^0.7.1",
"clsx": "^2.1.1",
"llama-stack-client": "^0.2.14",
"llama-stack-client": "^0.2.15",
"lucide-react": "^0.510.0",
"next": "15.3.3",
"next-auth": "^4.24.11",