Vector store inference api (#598)

# What does this PR do?
Moves all the memory providers to use the inference API and improved the
memory tests to setup the inference stack correctly and use the
embedding models


## Test Plan
torchrun $CONDA_PREFIX/bin/pytest -v -s -k "meta_reference"
--inference-model="Llama3.2-3B-Instruct"
--embedding-model="sentence-transformers/all-MiniLM-L6-v2"
llama_stack/providers/tests/inference/test_embeddings.py --env
EMBEDDING_DIMENSION=384


pytest -v -s llama_stack/providers/tests/memory/test_memory.py
--providers="inference=together,memory=weaviate"
--embedding-model="togethercomputer/m2-bert-80M-2k-retrieval" --env
EMBEDDING_DIMENSION=768 --env TOGETHER_API_KEY=<API-KEY> --env
WEAVIATE_API_KEY=foo --env WEAVIATE_CLUSTER_URL=bar
 
pytest -v -s llama_stack/providers/tests/memory/test_memory.py
--providers="inference=together,memory=chroma"
--embedding-model="togethercomputer/m2-bert-80M-2k-retrieval" --env
EMBEDDING_DIMENSION=768 --env TOGETHER_API_KEY=<API-KEY>--env
CHROMA_HOST=localhost --env CHROMA_PORT=8000

pytest -v -s llama_stack/providers/tests/memory/test_memory.py
--providers="inference=together,memory=pgvector"
--embedding-model="togethercomputer/m2-bert-80M-2k-retrieval" --env
PGVECTOR_DB=postgres --env PGVECTOR_USER=postgres --env
PGVECTOR_PASSWORD=mysecretpassword --env PGVECTOR_HOST=0.0.0.0 --env
EMBEDDING_DIMENSION=768 --env TOGETHER_API_KEY=<API-KEY>

pytest -v -s llama_stack/providers/tests/memory/test_memory.py
--providers="inference=together,memory=faiss"
--embedding-model="togethercomputer/m2-bert-80M-2k-retrieval" --env
EMBEDDING_DIMENSION=768 --env TOGETHER_API_KEY=<API-KEY>

llama_stack/providers/remote/memory/chroma/__init__.py

@@ -4,12 +4,15 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from typing import Dict
+
 from llama_stack.distribution.datatypes import RemoteProviderConfig
+from llama_stack.providers.datatypes import Api, ProviderSpec
 
 
-async def get_adapter_impl(config: RemoteProviderConfig, _deps):
+async def get_adapter_impl(config: RemoteProviderConfig, deps: Dict[Api, ProviderSpec]):
     from .chroma import ChromaMemoryAdapter
 
-    impl = ChromaMemoryAdapter(config.url)
+    impl = ChromaMemoryAdapter(config.url, deps[Api.inference])
     await impl.initialize()
     return impl

llama_stack/providers/remote/memory/chroma/chroma.py

@@ -15,8 +15,7 @@ from numpy.typing import NDArray
 from pydantic import parse_obj_as
 
 from llama_stack.apis.memory import *  # noqa: F403
-
-from llama_stack.providers.datatypes import MemoryBanksProtocolPrivate
+from llama_stack.providers.datatypes import Api, MemoryBanksProtocolPrivate
 from llama_stack.providers.utils.memory.vector_store import (
     BankWithIndex,
     EmbeddingIndex,
@@ -72,7 +71,7 @@ class ChromaIndex(EmbeddingIndex):
 
 
 class ChromaMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
-    def __init__(self, url: str) -> None:
+    def __init__(self, url: str, inference_api: Api.inference) -> None:
         log.info(f"Initializing ChromaMemoryAdapter with url: {url}")
         url = url.rstrip("/")
         parsed = urlparse(url)
@@ -82,6 +81,7 @@ class ChromaMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
 
         self.host = parsed.hostname
         self.port = parsed.port
+        self.inference_api = inference_api
 
         self.client = None
         self.cache = {}
@@ -109,10 +109,9 @@ class ChromaMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
             name=memory_bank.identifier,
             metadata={"bank": memory_bank.model_dump_json()},
         )
-        bank_index = BankWithIndex(
-            bank=memory_bank, index=ChromaIndex(self.client, collection)
+        self.cache[memory_bank.identifier] = BankWithIndex(
+            memory_bank, ChromaIndex(self.client, collection), self.inference_api
         )
-        self.cache[memory_bank.identifier] = bank_index
 
     async def list_memory_banks(self) -> List[MemoryBank]:
         collections = await self.client.list_collections()
@@ -124,11 +123,11 @@ class ChromaMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
                 log.exception(f"Failed to parse bank: {collection.metadata}")
                 continue
 
-            index = BankWithIndex(
-                bank=bank,
-                index=ChromaIndex(self.client, collection),
+            self.cache[bank.identifier] = BankWithIndex(
+                bank,
+                ChromaIndex(self.client, collection),
+                self.inference_api,
             )
-            self.cache[bank.identifier] = index
 
         return [i.bank for i in self.cache.values()]
 
@@ -166,6 +165,8 @@ class ChromaMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
         collection = await self.client.get_collection(bank_id)
         if not collection:
             raise ValueError(f"Bank {bank_id} not found in Chroma")
-        index = BankWithIndex(bank=bank, index=ChromaIndex(self.client, collection))
+        index = BankWithIndex(
+            bank, ChromaIndex(self.client, collection), self.inference_api
+        )
         self.cache[bank_id] = index
         return index

llama_stack/providers/remote/memory/pgvector/__init__.py

@@ -4,12 +4,16 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from typing import Dict
+
+from llama_stack.providers.datatypes import Api, ProviderSpec
+
 from .config import PGVectorConfig
 
 
-async def get_adapter_impl(config: PGVectorConfig, _deps):
+async def get_adapter_impl(config: PGVectorConfig, deps: Dict[Api, ProviderSpec]):
     from .pgvector import PGVectorMemoryAdapter
 
-    impl = PGVectorMemoryAdapter(config)
+    impl = PGVectorMemoryAdapter(config, deps[Api.inference])
     await impl.initialize()
     return impl

llama_stack/providers/remote/memory/pgvector/pgvector.py

@@ -16,9 +16,9 @@ from pydantic import BaseModel, parse_obj_as
 
 from llama_stack.apis.memory import *  # noqa: F403
 
-from llama_stack.providers.datatypes import MemoryBanksProtocolPrivate
+from llama_stack.providers.datatypes import Api, MemoryBanksProtocolPrivate
+
 from llama_stack.providers.utils.memory.vector_store import (
-    ALL_MINILM_L6_V2_DIMENSION,
     BankWithIndex,
     EmbeddingIndex,
 )
@@ -120,8 +120,9 @@ class PGVectorIndex(EmbeddingIndex):
 
 
 class PGVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
-    def __init__(self, config: PGVectorConfig) -> None:
+    def __init__(self, config: PGVectorConfig, inference_api: Api.inference) -> None:
         self.config = config
+        self.inference_api = inference_api
         self.cursor = None
         self.conn = None
         self.cache = {}
@@ -160,27 +161,17 @@ class PGVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
     async def shutdown(self) -> None:
         pass
 
-    async def register_memory_bank(
-        self,
-        memory_bank: MemoryBank,
-    ) -> None:
+    async def register_memory_bank(self, memory_bank: MemoryBank) -> None:
         assert (
             memory_bank.memory_bank_type == MemoryBankType.vector.value
         ), f"Only vector banks are supported {memory_bank.memory_bank_type}"
 
-        upsert_models(
-            self.cursor,
-            [
-                (memory_bank.identifier, memory_bank),
-            ],
+        upsert_models(self.cursor, [(memory_bank.identifier, memory_bank)])
+        index = PGVectorIndex(memory_bank, memory_bank.embedding_dimension, self.cursor)
+        self.cache[memory_bank.identifier] = BankWithIndex(
+            memory_bank, index, self.inference_api
         )
 
-        index = BankWithIndex(
-            bank=memory_bank,
-            index=PGVectorIndex(memory_bank, ALL_MINILM_L6_V2_DIMENSION, self.cursor),
-        )
-        self.cache[memory_bank.identifier] = index
-
     async def unregister_memory_bank(self, memory_bank_id: str) -> None:
         await self.cache[memory_bank_id].index.delete()
         del self.cache[memory_bank_id]
@@ -190,8 +181,9 @@ class PGVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
         for bank in banks:
             if bank.identifier not in self.cache:
                 index = BankWithIndex(
-                    bank=bank,
-                    index=PGVectorIndex(bank, ALL_MINILM_L6_V2_DIMENSION, self.cursor),
+                    bank,
+                    PGVectorIndex(bank, bank.embedding_dimension, self.cursor),
+                    self.inference_api,
                 )
                 self.cache[bank.identifier] = index
         return banks
@@ -214,14 +206,13 @@ class PGVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
         index = await self._get_and_cache_bank_index(bank_id)
         return await index.query_documents(query, params)
 
+        self.inference_api = inference_api
+
     async def _get_and_cache_bank_index(self, bank_id: str) -> BankWithIndex:
         if bank_id in self.cache:
             return self.cache[bank_id]
 
         bank = await self.memory_bank_store.get_memory_bank(bank_id)
-        index = BankWithIndex(
-            bank=bank,
-            index=PGVectorIndex(bank, ALL_MINILM_L6_V2_DIMENSION, self.cursor),
-        )
-        self.cache[bank_id] = index
-        return index
+        index = PGVectorIndex(bank, bank.embedding_dimension, self.cursor)
+        self.cache[bank_id] = BankWithIndex(bank, index, self.inference_api)
+        return self.cache[bank_id]

llama_stack/providers/remote/memory/qdrant/__init__.py

@@ -4,12 +4,16 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from typing import Dict
+
+from llama_stack.providers.datatypes import Api, ProviderSpec
+
 from .config import QdrantConfig
 
 
-async def get_adapter_impl(config: QdrantConfig, _deps):
+async def get_adapter_impl(config: QdrantConfig, deps: Dict[Api, ProviderSpec]):
     from .qdrant import QdrantVectorMemoryAdapter
 
-    impl = QdrantVectorMemoryAdapter(config)
+    impl = QdrantVectorMemoryAdapter(config, deps[Api.inference])
     await impl.initialize()
     return impl

llama_stack/providers/remote/memory/qdrant/qdrant.py

@@ -101,10 +101,11 @@ class QdrantIndex(EmbeddingIndex):
 
 
 class QdrantVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
-    def __init__(self, config: QdrantConfig) -> None:
+    def __init__(self, config: QdrantConfig, inference_api: Api.inference) -> None:
         self.config = config
         self.client = AsyncQdrantClient(**self.config.model_dump(exclude_none=True))
         self.cache = {}
+        self.inference_api = inference_api
 
     async def initialize(self) -> None:
         pass
@@ -123,6 +124,7 @@ class QdrantVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
         index = BankWithIndex(
             bank=memory_bank,
             index=QdrantIndex(self.client, memory_bank.identifier),
+            inference_api=self.inference_api,
         )
 
         self.cache[memory_bank.identifier] = index
@@ -143,6 +145,7 @@ class QdrantVectorMemoryAdapter(Memory, MemoryBanksProtocolPrivate):
         index = BankWithIndex(
             bank=bank,
             index=QdrantIndex(client=self.client, collection_name=bank_id),
+            inference_api=self.inference_api,
         )
         self.cache[bank_id] = index
         return index

llama_stack/providers/remote/memory/weaviate/__init__.py

@@ -4,12 +4,16 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from typing import Dict
+
+from llama_stack.providers.datatypes import Api, ProviderSpec
+
 from .config import WeaviateConfig, WeaviateRequestProviderData  # noqa: F401
 
 
-async def get_adapter_impl(config: WeaviateConfig, _deps):
+async def get_adapter_impl(config: WeaviateConfig, deps: Dict[Api, ProviderSpec]):
     from .weaviate import WeaviateMemoryAdapter
 
-    impl = WeaviateMemoryAdapter(config)
+    impl = WeaviateMemoryAdapter(config, deps[Api.inference])
     await impl.initialize()
     return impl

llama_stack/providers/remote/memory/weaviate/weaviate.py

@@ -12,10 +12,11 @@ import weaviate
 import weaviate.classes as wvc
 from numpy.typing import NDArray
 from weaviate.classes.init import Auth
+from weaviate.classes.query import Filter
 
 from llama_stack.apis.memory import *  # noqa: F403
 from llama_stack.distribution.request_headers import NeedsRequestProviderData
-from llama_stack.providers.datatypes import MemoryBanksProtocolPrivate
+from llama_stack.providers.datatypes import Api, MemoryBanksProtocolPrivate
 from llama_stack.providers.utils.memory.vector_store import (
     BankWithIndex,
     EmbeddingIndex,
@@ -80,12 +81,21 @@ class WeaviateIndex(EmbeddingIndex):
 
         return QueryDocumentsResponse(chunks=chunks, scores=scores)
 
+    async def delete(self, chunk_ids: List[str]) -> None:
+        collection = self.client.collections.get(self.collection_name)
+        collection.data.delete_many(
+            where=Filter.by_property("id").contains_any(chunk_ids)
+        )
+
 
 class WeaviateMemoryAdapter(
-    Memory, NeedsRequestProviderData, MemoryBanksProtocolPrivate
+    Memory,
+    NeedsRequestProviderData,
+    MemoryBanksProtocolPrivate,
 ):
-    def __init__(self, config: WeaviateConfig) -> None:
+    def __init__(self, config: WeaviateConfig, inference_api: Api.inference) -> None:
         self.config = config
+        self.inference_api = inference_api
         self.client_cache = {}
         self.cache = {}
 
@@ -117,7 +127,7 @@ class WeaviateMemoryAdapter(
         memory_bank: MemoryBank,
     ) -> None:
         assert (
-            memory_bank.memory_bank_type == MemoryBankType.vector
+            memory_bank.memory_bank_type == MemoryBankType.vector.value
         ), f"Only vector banks are supported {memory_bank.memory_bank_type}"
 
         client = self._get_client()
@@ -135,11 +145,11 @@ class WeaviateMemoryAdapter(
                 ],
             )
 
-        index = BankWithIndex(
-            bank=memory_bank,
-            index=WeaviateIndex(client=client, collection_name=memory_bank.identifier),
+        self.cache[memory_bank.identifier] = BankWithIndex(
+            memory_bank,
+            WeaviateIndex(client=client, collection_name=memory_bank.identifier),
+            self.inference_api,
         )
-        self.cache[memory_bank.identifier] = index
 
     async def list_memory_banks(self) -> List[MemoryBank]:
         # TODO: right now the Llama Stack is the source of truth for these banks. That is
@@ -163,6 +173,7 @@ class WeaviateMemoryAdapter(
         index = BankWithIndex(
             bank=bank,
             index=WeaviateIndex(client=client, collection_name=bank_id),
+            inference_api=self.inference_api,
         )
         self.cache[bank_id] = index
         return index