add ltm

Author: yaozheng-fang

veadk/knowledgebase/backends/base_backend.py

@@ -24,31 +24,29 @@ class BaseKnowledgebaseBackend(ABC, BaseModel):
     @abstractmethod
     def add_from_directory(self, directory: str, **kwargs) -> bool:
         """Add knowledge from file path to knowledgebase"""
-        ...
 
     @abstractmethod
     def add_from_files(self, files: list[str], **kwargs) -> bool:
         """Add knowledge (e.g, documents, strings, ...) to knowledgebase"""
-        ...
 
     @abstractmethod
     def add_from_text(self, text: str | list[str], **kwargs) -> bool:
         """Add knowledge from text to knowledgebase"""
-        ...
 
     @abstractmethod
     def search(self, **kwargs) -> list:
         """Search knowledge from knowledgebase"""
-        ...
 
-    def delete(self, **kwargs) -> bool:
-        """Delete knowledge from knowledgebase"""
-        ...
+    # Optional methods for future use:
+    # - `delete`: Delete collection or documents
+    # - `list_docs`: List original documents
+    # - `list_chunks`: List embedded document chunks
 
-    def list_docs(self, **kwargs) -> None:
-        """List original documents in knowledgebase"""
-        pass
+    # def delete(self, **kwargs) -> bool:
+    #     """Delete knowledge from knowledgebase"""
 
-    def list_chunks(self, **kwargs) -> None:
-        """List embeded document chunks in knowledgebase"""
-        pass
+    # def list_docs(self, **kwargs) -> None:
+    #     """List original documents in knowledgebase"""
+
+    # def list_chunks(self, **kwargs) -> None:
+    #     """List embeded document chunks in knowledgebase"""

veadk/knowledgebase/backends/redis_backend.py

@@ -21,6 +21,10 @@
 from llama_index.core.schema import BaseNode
 from llama_index.embeddings.openai_like import OpenAILikeEmbedding
 from llama_index.vector_stores.redis import RedisVectorStore
+from llama_index.vector_stores.redis.schema import (
+    RedisIndexInfo,
+    RedisVectorStoreSchema,
+)
 from pydantic import Field
 from redis import Redis
 from typing_extensions import Any, override
@@ -56,8 +60,14 @@ def model_post_init(self, __context: Any) -> None:
             api_base=self.embedding_config.api_base,
         )
 
+        self._schema = RedisVectorStoreSchema(
+            index=RedisIndexInfo(name=self.index),
+        )
         self._vector_store = RedisVectorStore(
-            redis_client=self._redis_client, overwrite=True
+            schema=self._schema,
+            redis_client=self._redis_client,
+            overwrite=True,
+            collection_name=self.index,
         )
 
         self._storage_context = StorageContext.from_defaults(

veadk/memory/long_term_memory.py

@@ -25,45 +25,83 @@
 from google.adk.memory.memory_entry import MemoryEntry
 from google.adk.sessions import Session
 from google.genai import types
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 from typing_extensions import override
 
-from veadk.database import DatabaseFactory
-from veadk.database.database_adapter import get_long_term_memory_database_adapter
+from veadk.memory.long_term_memory_backends.base_backend import (
+    BaseLongTermMemoryBackend,
+)
+from veadk.memory.long_term_memory_backends.in_memory_backend import InMemoryLTMBackend
+from veadk.memory.long_term_memory_backends.opensearch_backend import (
+    OpensearchLTMBackend,
+)
+from veadk.memory.long_term_memory_backends.redis_backend import RedisLTMBackend
+from veadk.memory.long_term_memory_backends.vikingdb_memory_backend import (
+    VikingDBKnowledgeBackend,
+)
 from veadk.utils.logger import get_logger
 
 logger = get_logger(__name__)
 
 
+BACKEND_CLS = {
+    "local": InMemoryLTMBackend,
+    "opensearch": OpensearchLTMBackend,
+    "viking": VikingDBKnowledgeBackend,
+    "viking_mem": VikingDBKnowledgeBackend,
+    "redis": RedisLTMBackend,
+}
+
+
 def build_long_term_memory_index(app_name: str, user_id: str):
     return f"{app_name}_{user_id}"
 
 
 class LongTermMemory(BaseMemoryService, BaseModel):
-    backend: Literal[
-        "local", "opensearch", "redis", "mysql", "viking", "viking_mem"
-    ] = "opensearch"
+    backend: Literal["local", "opensearch", "redis", "viking", "viking_mem"] = (
+        "opensearch"
+    )
+    """Long term memory backend type"""
+
+    backend_config: dict = Field(default_factory=dict)
+    """Long term memory backend configuration"""
+
+    backend_instance: BaseLongTermMemoryBackend | None = None
+    """An instance of a long term memory backend that implements the `BaseLongTermMemoryBackend` interface."""
+
     top_k: int = 5
+    """Number of top similar documents to retrieve during search."""
 
-    def model_post_init(self, __context: Any) -> None:
-        if self.backend == "viking":
-            logger.warning(
-                "`viking` backend is deprecated, switching to `viking_mem` backend."
-            )
-            self.backend = "viking_mem"
+    app_name: str = ""
 
-        logger.info(
-            f"Initializing long term memory: backend={self.backend} top_k={self.top_k}"
-        )
+    user_id: str = ""
 
-        self._db_client = DatabaseFactory.create(
-            backend=self.backend,
-        )
-        self._adapter = get_long_term_memory_database_adapter(self._db_client)
+    def model_post_init(self, __context: Any) -> None:
+        self._backend = None
 
-        logger.info(
-            f"Initialized long term memory: db_client={self._db_client.__class__.__name__} adapter={self._adapter}"
-        )
+        if self.backend_instance:
+            self._backend = self.backend_instance
+            logger.info(
+                f"Initialized long term memory with provided backend instance {self._backend.__class__.__name__}"
+            )
+        else:
+            if self.backend_config:
+                logger.info(
+                    f"Initialized long term memory backend {self.backend} with config."
+                )
+                self._backend = BACKEND_CLS[self.backend](**self.backend_config)
+            elif self.app_name and self.user_id:
+                self.index = build_long_term_memory_index(
+                    app_name=self.app_name, user_id=self.user_id
+                )
+                logger.info(f"Long term memory index set to {self.index}.")
+                self._backend = BACKEND_CLS[self.backend](
+                    **self.backend_config, index=self.index
+                )
+            else:
+                logger.warning(
+                    "Neither `backend_instance`, `backend_config`, nor `app_name`/`user_id` is provided, the long term memory storage will initialize when adding a session."
+                )
 
     def _filter_and_convert_events(self, events: list[Event]) -> list[str]:
         final_events = []
@@ -91,40 +129,48 @@ async def add_session_to_memory(
         self,
         session: Session,
     ):
+        app_name = session.app_name
+        user_id = session.user_id
+
+        if self.index != build_long_term_memory_index(app_name, user_id):
+            logger.warning(
+                f"The `app_name` or `user_id` is different from the initialized one, skip add session to memory. Initialized index: {self.index}, current built index: {build_long_term_memory_index(app_name, user_id)}"
+            )
+            return
+
+        if not self._backend:
+            self.index = build_long_term_memory_index(app_name, user_id)
+            self._backend = BACKEND_CLS[self.backend](index=self.index)
+            logger.info(
+                f"Initialize long term memory backend now, index is {self.index}"
+            )
+
         event_strings = self._filter_and_convert_events(session.events)
-        index = build_long_term_memory_index(session.app_name, session.user_id)
 
         logger.info(
-            f"Adding {len(event_strings)} events to long term memory: index={index}"
+            f"Adding {len(event_strings)} events to long term memory: index={self.index}"
         )
 
-        # check if viking memory database, should give a user id： if/else
-        if self.backend == "viking_mem":
-            self._adapter.add(data=event_strings, index=index, user_id=session.user_id)
-        else:
-            self._adapter.add(data=event_strings, index=index)
+        self._backend.save_memory(event_strings=event_strings)
 
         logger.info(
-            f"Added {len(event_strings)} events to long term memory: index={index}"
+            f"Added {len(event_strings)} events to long term memory: index={self.index}"
         )
 
     @override
     async def search_memory(self, *, app_name: str, user_id: str, query: str):
-        index = build_long_term_memory_index(app_name, user_id)
-
         logger.info(
-            f"Searching long term memory: query={query} index={index} top_k={self.top_k}"
+            f"Searching long term memory: query={query} index={self.index} top_k={self.top_k}"
         )
 
-        # user id if viking memory db
-        if self.backend == "viking_mem":
-            memory_chunks = self._adapter.query(
-                query=query, index=index, top_k=self.top_k, user_id=user_id
-            )
-        else:
-            memory_chunks = self._adapter.query(
-                query=query, index=index, top_k=self.top_k
+        # prevent model invoke `load_memory` before add session to this memory
+        if not self._backend:
+            logger.error(
+                "Long term memory backend is not initialized, cannot search memory."
             )
+            return SearchMemoryResponse(memories=[])
+
+        memory_chunks = self._backend.search_memory(query=query, top_k=self.top_k)
 
         memory_events = []
         for memory in memory_chunks:
@@ -152,6 +198,6 @@ async def search_memory(self, *, app_name: str, user_id: str, query: str):
             )
 
         logger.info(
-            f"Return {len(memory_events)} memory events for query: {query} index={index}"
+            f"Return {len(memory_events)} memory events for query: {query} index={self.index}"
         )
         return SearchMemoryResponse(memories=memory_events)

veadk/memory/long_term_memory_backends/base_backend.py

@@ -0,0 +1,29 @@
+# Copyright (c) 2025 Beijing Volcano Engine Technology Co., Ltd. and/or its affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import ABC, abstractmethod
+
+from pydantic import BaseModel
+
+
+class BaseLongTermMemoryBackend(ABC, BaseModel):
+    index: str
+
+    @abstractmethod
+    def save_memory(self, event_strings: list[str], **kwargs) -> bool:
+        """Save memory to long term memory backend"""
+
+    @abstractmethod
+    def search_memory(self, query: str, **kwargs) -> list[str]:
+        """Retrieve memory from long term memory backend"""

veadk/memory/long_term_memory_backends/in_memory_backend.py

@@ -0,0 +1,61 @@
+# Copyright (c) 2025 Beijing Volcano Engine Technology Co., Ltd. and/or its affiliates.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from llama_index.core import Document, VectorStoreIndex
+from llama_index.core.schema import BaseNode
+from llama_index.embeddings.openai_like import OpenAILikeEmbedding
+from pydantic import Field
+from typing_extensions import Any, override
+
+from veadk.configs.model_configs import EmbeddingModelConfig
+from veadk.knowledgebase.backends.utils import get_llama_index_splitter
+from veadk.memory.long_term_memory_backends.base_backend import (
+    BaseLongTermMemoryBackend,
+)
+
+
+class InMemoryLTMBackend(BaseLongTermMemoryBackend):
+    embedding_config: EmbeddingModelConfig = Field(default_factory=EmbeddingModelConfig)
+    """Embedding model configs"""
+
+    def model_post_init(self, __context: Any) -> None:
+        self._embed_model = OpenAILikeEmbedding(
+            model_name=self.embedding_config.name,
+            api_key=self.embedding_config.api_key,
+            api_base=self.embedding_config.api_base,
+        )
+        self._vector_index = VectorStoreIndex([], embed_model=self._embed_model)
+        self._retriever = self._vector_index.as_retriever()
+
+    @override
+    def save_memory(self, event_strings: list[str], **kwargs) -> bool:
+        for event_string in event_strings:
+            document = Document(text=event_string)
+            nodes = self._split_documents([document])
+            self._vector_index.insert_nodes(nodes)
+        return True
+
+    @override
+    def search_memory(self, query: str, top_k: int, **kwargs) -> list[str]:
+        retrieved_nodes = self._retriever.retrieve(query, top_k=top_k)
+        return [node.text for node in retrieved_nodes]
+
+    def _split_documents(self, documents: list[Document]) -> list[BaseNode]:
+        """Split document into chunks"""
+        nodes = []
+        for document in documents:
+            splitter = get_llama_index_splitter(document.metadata.get("file_path", ""))
+            _nodes = splitter.get_nodes_from_documents([document])
+            nodes.extend(_nodes)
+        return nodes