flesh out memory banks API

llama_toolchain/agentic_system/api/datatypes.py

@@ -10,13 +10,13 @@ from typing import Any, Dict, List, Literal, Optional, Union
 
 from llama_models.schema_utils import json_schema_type
 
-from pydantic import BaseModel, ConfigDict, Field, validator
+from pydantic import BaseModel, ConfigDict, Field
 from typing_extensions import Annotated
 
 from llama_toolchain.common.deployment_types import *  # noqa: F403
 from llama_toolchain.inference.api import *  # noqa: F403
-from llama_toolchain.safety.api.datatypes import *  # noqa: F403
-from llama_toolchain.memory.api.datatypes import *  # noqa: F403
+from llama_toolchain.safety.api import *  # noqa: F403
+from llama_toolchain.memory.api import *  # noqa: F403
 
 
 @json_schema_type
@@ -25,31 +25,81 @@ class Attachment(BaseModel):
     mime_type: str
 
 
-class AgenticSystemBuiltinTool(BuiltinTool):
+class AgenticSystemTool(Enum):
+    brave_search = "brave_search"
+    wolfram_alpha = "wolfram_alpha"
+    photogen = "photogen"
+    code_interpreter = "code_interpreter"
+
+    function_call = "function_call"
     memory = "memory"
 
 
-@json_schema_type
-class AgenticSystemToolDefinition(BaseModel):
-    tool_name: Union[AgenticSystemBuiltinTool, str]
-    description: Optional[str] = None
-    parameters: Optional[Dict[str, ToolParamDefinition]] = None
-
-    @validator("tool_name", pre=True)
-    @classmethod
-    def validate_field(cls, v):
-        if isinstance(v, str):
-            try:
-                return AgenticSystemBuiltinTool(v)
-            except ValueError:
-                return v
-        return v
-
-    execution_config: Optional[RestAPIExecutionConfig] = None
+class ToolDefinitionCommon(BaseModel):
     input_shields: Optional[List[ShieldDefinition]] = Field(default_factory=list)
     output_shields: Optional[List[ShieldDefinition]] = Field(default_factory=list)
 
 
+@json_schema_type
+class BraveSearchToolDefinition(ToolDefinitionCommon):
+    type: Literal[AgenticSystemTool.brave_search.value] = (
+        AgenticSystemTool.brave_search.value
+    )
+    remote_execution: Optional[RestAPIExecutionConfig] = None
+
+
+@json_schema_type
+class WolframAlphaToolDefinition(ToolDefinitionCommon):
+    type: Literal[AgenticSystemTool.wolfram_alpha.value] = (
+        AgenticSystemTool.wolfram_alpha.value
+    )
+    remote_execution: Optional[RestAPIExecutionConfig] = None
+
+
+@json_schema_type
+class PhotogenToolDefinition(ToolDefinitionCommon):
+    type: Literal[AgenticSystemTool.photogen.value] = AgenticSystemTool.photogen.value
+    remote_execution: Optional[RestAPIExecutionConfig] = None
+
+
+@json_schema_type
+class CodeInterpreterToolDefinition(ToolDefinitionCommon):
+    type: Literal[AgenticSystemTool.code_interpreter.value] = (
+        AgenticSystemTool.code_interpreter.value
+    )
+    enable_inline_code_execution: bool = True
+    remote_execution: Optional[RestAPIExecutionConfig] = None
+
+
+@json_schema_type
+class FunctionCallToolDefinition(ToolDefinitionCommon):
+    type: Literal[AgenticSystemTool.function_call.value] = (
+        AgenticSystemTool.function_call.value
+    )
+    description: str
+    parameters: Dict[str, ToolParamDefinition]
+    remote_execution: Optional[RestAPIExecutionConfig] = None
+
+
+@json_schema_type
+class MemoryToolDefinition(ToolDefinitionCommon):
+    type: Literal[AgenticSystemTool.memory.value] = AgenticSystemTool.memory.value
+    memory_banks: List[MemoryBank] = Field(default_factory=list)
+
+
+AgenticSystemToolDefinition = Annotated[
+    Union[
+        BraveSearchToolDefinition,
+        WolframAlphaToolDefinition,
+        PhotogenToolDefinition,
+        CodeInterpreterToolDefinition,
+        FunctionCallToolDefinition,
+        MemoryToolDefinition,
+    ],
+    Field(discriminator="type"),
+]
+
+
 class StepCommon(BaseModel):
     turn_id: str
     step_id: str
@@ -136,27 +186,45 @@ class Session(BaseModel):
     started_at: datetime
 
 
-@json_schema_type
-class MemoryConfig(BaseModel):
-    memory_bank_id: str
+class MemoryBankConfigCommon(BaseModel):
+    bank_id: str
 
-    # this configuration can hold other information we may want to configure
-    # how will the agent use the memory bank API?
-    #
-    #
+
+class VectorMemoryBankConfig(MemoryBankConfigCommon):
+    type: Literal[MemoryBankType.vector.value] = MemoryBankType.vector.value
+
+
+class KeyValueMemoryBankConfig(MemoryBankConfigCommon):
+    type: Literal[MemoryBankType.keyvalue.value] = MemoryBankType.keyvalue.value
+    keys: List[str]  # what keys to focus on
+
+
+class KeywordMemoryBankConfig(MemoryBankConfigCommon):
+    type: Literal[MemoryBankType.keyword.value] = MemoryBankType.keyword.value
+
+
+class GraphMemoryBankConfig(MemoryBankConfigCommon):
+    type: Literal[MemoryBankType.graph.value] = MemoryBankType.graph.value
+    entities: List[str]  # what entities to focus on
+
+
+MemoryBankConfig = Annotated[
+    Union[
+        VectorMemoryBankConfig,
+        KeyValueMemoryBankConfig,
+        KeywordMemoryBankConfig,
+        GraphMemoryBankConfig,
+    ],
+    Field(discriminator="type"),
+]
 
 
 class AgentConfigCommon(BaseModel):
     sampling_params: Optional[SamplingParams] = SamplingParams()
-    memory_configs: Optional[List[MemoryConfig]] = Field(default_factory=list)
 
     input_shields: Optional[List[ShieldDefinition]] = Field(default_factory=list)
     output_shields: Optional[List[ShieldDefinition]] = Field(default_factory=list)
 
-    # if you completely want to replace the messages prefixed by the system,
-    # this is debug only
-    debug_prefix_messages: Optional[List[Message]] = Field(default_factory=list)
-
     tools: Optional[List[AgenticSystemToolDefinition]] = Field(default_factory=list)
     tool_choice: Optional[ToolChoice] = Field(default=ToolChoice.auto)
     tool_prompt_format: Optional[ToolPromptFormat] = Field(
@@ -168,6 +236,7 @@ class AgentConfigCommon(BaseModel):
 class AgentConfig(AgentConfigCommon):
     model: str
     instructions: str
+    memory_bank_configs: Optional[List[MemoryBankConfig]] = Field(default_factory=list)
 
 
 class AgentConfigOverridablePerTurn(AgentConfigCommon):

llama_toolchain/agentic_system/api/endpoints.py

@@ -10,21 +10,9 @@ from typing import Protocol
 from llama_models.schema_utils import json_schema_type, webmethod
 
 
-@json_schema_type
-class AgenticSystemCreateRequest(BaseModel):
-    agent_config: AgentConfig
-
-
 @json_schema_type
 class AgenticSystemCreateResponse(BaseModel):
-    # TODO: rename this to agent_id
-    system_id: str
-
-
-@json_schema_type
-class AgenticSystemSessionCreateRequest(BaseModel):
-    system_id: str
-    session_name: str
+    agent_id: str
 
 
 @json_schema_type
@@ -33,8 +21,8 @@ class AgenticSystemSessionCreateResponse(BaseModel):
 
 
 @json_schema_type
-class AgenticSystemTurnCreateRequest(BaseModel, AgentConfigOverridablePerTurn):
-    system_id: str
+class AgenticSystemTurnCreateRequest(AgentConfigOverridablePerTurn):
+    agent_id: str
     session_id: str
 
     # TODO: figure out how we can simplify this and make why
@@ -67,7 +55,7 @@ class AgenticSystem(Protocol):
     @webmethod(route="/agentic_system/create")
     async def create_agentic_system(
         self,
-        request: AgenticSystemCreateRequest,
+        agent_config: AgentConfig,
     ) -> AgenticSystemCreateResponse: ...
 
     @webmethod(route="/agentic_system/turn/create")
@@ -91,7 +79,8 @@ class AgenticSystem(Protocol):
     @webmethod(route="/agentic_system/session/create")
     async def create_agentic_system_session(
         self,
-        request: AgenticSystemSessionCreateRequest,
+        agent_id: str,
+        session_name: str,
     ) -> AgenticSystemSessionCreateResponse: ...
 
     @webmethod(route="/agentic_system/session/get")

llama_toolchain/agentic_system/client.py

@@ -26,9 +26,7 @@ from llama_toolchain.agentic_system.event_logger import EventLogger
 from .api import (
     AgentConfig,
     AgenticSystem,
-    AgenticSystemCreateRequest,
     AgenticSystemCreateResponse,
-    AgenticSystemSessionCreateRequest,
     AgenticSystemSessionCreateResponse,
     AgenticSystemToolDefinition,
     AgenticSystemTurnCreateRequest,
@@ -127,28 +125,24 @@ async def run_main(host: str, port: int):
         ),
     ]
 
-    create_request = AgenticSystemCreateRequest(
+    agent_config = AgentConfig(
         model="Meta-Llama3.1-8B-Instruct",
-        agent_config=AgentConfig(
         instructions="You are a helpful assistant",
         sampling_params=SamplingParams(),
-            available_tools=tool_definitions,
+        tools=tool_definitions,
         input_shields=[],
         output_shields=[],
         debug_prefix_messages=[],
         tool_prompt_format=ToolPromptFormat.json,
-        ),
     )
 
-    create_response = await api.create_agentic_system(create_request)
+    create_response = await api.create_agentic_system(agent_config)
     print(create_response)
 
     session_response = await api.create_agentic_system_session(
-        AgenticSystemSessionCreateRequest(
-            system_id=create_response.system_id,
+        agent_id=create_response.agent_id,
         session_name="test_session",
     )
-    )
     print(session_response)
 
     user_prompts = [
@@ -162,7 +156,7 @@ async def run_main(host: str, port: int):
         cprint(f"User> {content}", color="blue")
         iterator = api.create_agentic_system_turn(
             AgenticSystemTurnCreateRequest(
-                system_id=create_response.system_id,
+                agent_id=create_response.agent_id,
                 session_id=session_response.session_id,
                 messages=[
                     UserMessage(content=content),

llama_toolchain/agentic_system/meta_reference/agent_instance.py

@@ -8,7 +8,7 @@
 import copy
 import uuid
 from datetime import datetime
-from typing import AsyncGenerator, List, Optional
+from typing import AsyncGenerator, List
 
 from llama_models.llama3.api.datatypes import ToolPromptFormat
 
@@ -326,7 +326,7 @@ class ChatAgent(ShieldRunnerMixin):
             req = ChatCompletionRequest(
                 model=self.agent_config.model,
                 messages=input_messages,
-                tools=self.agent_config.available_tools,
+                tools=self.agent_config.tools,
                 tool_prompt_format=self.agent_config.tool_prompt_format,
                 stream=True,
                 sampling_params=sampling_params,

llama_toolchain/agentic_system/meta_reference/agentic_system.py

@@ -24,7 +24,7 @@ from llama_toolchain.tools.builtin import (
 )
 from llama_toolchain.tools.safety import with_safety
 
-from .agent_instance import AgentInstance, ChatAgent
+from .agent_instance import ChatAgent
 from .config import MetaReferenceImplConfig
 
 
@@ -71,11 +71,11 @@ class MetaReferenceAgenticSystemImpl(AgenticSystem):
         self,
         request: AgenticSystemCreateRequest,
     ) -> AgenticSystemCreateResponse:
-        system_id = str(uuid.uuid4())
+        agent_id = str(uuid.uuid4())
 
         builtin_tools = []
         cfg = request.agent_config
-        for dfn in cfg.available_tools:
+        for dfn in cfg.tools:
             if isinstance(dfn.tool_name, BuiltinTool):
                 if dfn.tool_name == BuiltinTool.wolfram_alpha:
                     key = self.config.wolfram_api_key
@@ -102,7 +102,7 @@ class MetaReferenceAgenticSystemImpl(AgenticSystem):
                     )
                 )
 
-        AGENT_INSTANCES_BY_ID[system_id] = ChatAgent(
+        AGENT_INSTANCES_BY_ID[agent_id] = ChatAgent(
             agent_config=cfg,
             inference_api=self.inference_api,
             safety_api=self.safety_api,
@@ -111,16 +111,16 @@ class MetaReferenceAgenticSystemImpl(AgenticSystem):
         )
 
         return AgenticSystemCreateResponse(
-            system_id=system_id,
+            agent_id=agent_id,
         )
 
     async def create_agentic_system_session(
         self,
         request: AgenticSystemSessionCreateRequest,
     ) -> AgenticSystemSessionCreateResponse:
-        system_id = request.system_id
-        assert system_id in AGENT_INSTANCES_BY_ID, f"System {system_id} not found"
-        agent = AGENT_INSTANCES_BY_ID[system_id]
+        agent_id = request.agent_id
+        assert agent_id in AGENT_INSTANCES_BY_ID, f"System {agent_id} not found"
+        agent = AGENT_INSTANCES_BY_ID[agent_id]
 
         session = agent.create_session(request.session_name)
         return AgenticSystemSessionCreateResponse(
@@ -131,9 +131,9 @@ class MetaReferenceAgenticSystemImpl(AgenticSystem):
         self,
         request: AgenticSystemTurnCreateRequest,
     ) -> AsyncGenerator:
-        system_id = request.system_id
-        assert system_id in AGENT_INSTANCES_BY_ID, f"System {system_id} not found"
-        agent = AGENT_INSTANCES_BY_ID[system_id]
+        agent_id = request.agent_id
+        assert agent_id in AGENT_INSTANCES_BY_ID, f"System {agent_id} not found"
+        agent = AGENT_INSTANCES_BY_ID[agent_id]
 
         assert (
             request.session_id in agent.sessions

llama_toolchain/agentic_system/meta_reference/execute_with_custom_tools.py

@@ -19,7 +19,7 @@ from llama_toolchain.inference.api import Message
 
 async def execute_with_custom_tools(
     system: AgenticSystem,
-    system_id: str,
+    agent_id: str,
     session_id: str,
     messages: List[Message],
     custom_tools: List[Any],
@@ -35,7 +35,7 @@ async def execute_with_custom_tools(
         n_iter += 1
 
         request = AgenticSystemTurnCreateRequest(
-            system_id=system_id,
+            agent_id=agent_id,
             session_id=session_id,
             messages=current_messages,
             stream=stream,

llama_toolchain/agentic_system/utils.py

@@ -14,12 +14,7 @@ from llama_models.llama3.api.datatypes import (
     ToolPromptFormat,
 )
 
-from llama_toolchain.agentic_system.api import (
-    AgentConfig,
-    AgenticSystemCreateRequest,
-    AgenticSystemSessionCreateRequest,
-    AgenticSystemToolDefinition,
-)
+from llama_toolchain.agentic_system.api import AgentConfig, AgenticSystemToolDefinition
 from llama_toolchain.agentic_system.client import AgenticSystemClient
 
 from llama_toolchain.agentic_system.meta_reference.execute_with_custom_tools import (
@@ -32,9 +27,9 @@ from llama_toolchain.safety.api.datatypes import BuiltinShield, ShieldDefinition
 
 
 class AgenticSystemClientWrapper:
-    def __init__(self, api, system_id, custom_tools):
+    def __init__(self, api, agent_id, custom_tools):
         self.api = api
-        self.system_id = system_id
+        self.agent_id = agent_id
         self.custom_tools = custom_tools
         self.session_id = None
 
@@ -43,18 +38,16 @@ class AgenticSystemClientWrapper:
             name = f"Session-{uuid.uuid4()}"
 
         response = await self.api.create_agentic_system_session(
-            AgenticSystemSessionCreateRequest(
-                system_id=self.system_id,
+            agent_id=self.agent_id,
             session_name=name,
         )
-        )
         self.session_id = response.session_id
         return self.session_id
 
     async def run(self, messages: List[Message], stream: bool = True):
         async for chunk in execute_with_custom_tools(
             self.api,
-            self.system_id,
+            self.agent_id,
             self.session_id,
             messages,
             self.custom_tools,
@@ -98,11 +91,10 @@ async def get_agent_system_instance(
                 ShieldDefinition(shield_type=BuiltinShield.injection_shield),
             ]
 
-    create_request = AgenticSystemCreateRequest(
+    agent_config = AgentConfig(
         model=model,
-        agent_config=AgentConfig(
         instructions="You are a helpful assistant",
-            available_tools=tool_definitions,
+        tools=tool_definitions,
         input_shields=(
             []
             if disable_safety
@@ -120,7 +112,6 @@ async def get_agent_system_instance(
         ),
         sampling_params=SamplingParams(),
         tool_prompt_format=tool_prompt_format,
-        ),
     )
-    create_response = await api.create_agentic_system(create_request)
-    return AgenticSystemClientWrapper(api, create_response.system_id, custom_tools)
+    create_response = await api.create_agentic_system(agent_config)
+    return AgenticSystemClientWrapper(api, create_response.agent_id, custom_tools)

llama_toolchain/memory/api/datatypes.py

@@ -3,23 +3,3 @@
 #
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
-
-from typing import Any, Dict
-
-from llama_models.schema_utils import json_schema_type
-
-from pydantic import BaseModel
-
-
-@json_schema_type
-class MemoryBank(BaseModel):
-    memory_bank_id: str
-    memory_bank_name: str
-
-
-@json_schema_type
-class MemoryBankDocument(BaseModel):
-    document_id: str
-    content: bytes
-    metadata: Dict[str, Any]
-    mime_type: str

llama_toolchain/memory/api/endpoints.py

@@ -6,76 +6,132 @@
 
 from typing import List, Protocol
 
-from llama_models.llama3.api.datatypes import InterleavedTextMedia
+from pydantic import BaseModel, Field
+from typing_extensions import Annotated
+
+from llama_models.llama3.api.datatypes import *  # noqa: F403
 
 from llama_models.schema_utils import webmethod
 from .datatypes import *  # noqa: F403
 
 
 @json_schema_type
-class RetrieveMemoryDocumentsRequest(BaseModel):
-    query: InterleavedTextMedia
-    bank_ids: str
+class MemoryBankDocument(BaseModel):
+    document_id: str
+    content: InterleavedTextMedia | URL
+    mime_type: str
+    metadata: Dict[str, Any]
+
+
+class Chunk(BaseModel):
+    content: InterleavedTextMedia
+    token_count: int
 
 
 @json_schema_type
-class RetrieveMemoryDocumentsResponse(BaseModel):
-    documents: List[MemoryBankDocument]
+class QueryDocumentsResponse(BaseModel):
+    chunks: List[Chunk]
     scores: List[float]
 
 
+@json_schema_type
+class MemoryBankType(Enum):
+    vector = "vector"
+    keyvalue = "keyvalue"
+    keyword = "keyword"
+    graph = "graph"
+
+
+class VectorMemoryBankConfig(BaseModel):
+    type: Literal[MemoryBankType.vector.value] = MemoryBankType.vector.value
+    embedding_model: str
+
+
+class KeyValueMemoryBankConfig(BaseModel):
+    type: Literal[MemoryBankType.keyvalue.value] = MemoryBankType.keyvalue.value
+
+
+class KeywordMemoryBankConfig(BaseModel):
+    type: Literal[MemoryBankType.keyword.value] = MemoryBankType.keyword.value
+
+
+class GraphMemoryBankConfig(BaseModel):
+    type: Literal[MemoryBankType.graph.value] = MemoryBankType.graph.value
+
+
+MemoryBankConfig = Annotated[
+    Union[
+        VectorMemoryBankConfig,
+        KeyValueMemoryBankConfig,
+        KeywordMemoryBankConfig,
+        GraphMemoryBankConfig,
+    ],
+    Field(discriminator="type"),
+]
+
+
+@json_schema_type
+class MemoryBank(BaseModel):
+    bank_id: str
+    name: str
+    config: MemoryBankConfig
+    # if there's a pre-existing store which obeys the MemoryBank REST interface
+    url: Optional[URL] = None
+
+
 class Memory(Protocol):
     @webmethod(route="/memory_banks/create")
     def create_memory_bank(
         self,
-        bank_id: str,
-        bank_name: str,
-        embedding_model: str,
-        documents: List[MemoryBankDocument],
-    ) -> None: ...
+        name: str,
+        config: MemoryBankConfig,
+        url: Optional[URL] = None,
+    ) -> MemoryBank: ...
 
-    @webmethod(route="/memory_banks/list")
-    def get_memory_banks(self) -> List[MemoryBank]: ...
+    @webmethod(route="/memory_banks/list", method="GET")
+    def list_memory_banks(self) -> List[MemoryBank]: ...
 
     @webmethod(route="/memory_banks/get")
-    def get_memory_bank(self, bank_id: str) -> List[MemoryBank]: ...
+    def get_memory_bank(self, bank_id: str) -> MemoryBank: ...
 
-    @webmethod(route="/memory_banks/drop")
-    def delete_memory_bank(
+    @webmethod(route="/memory_banks/drop", method="DELETE")
+    def drop_memory_bank(
         self,
         bank_id: str,
     ) -> str: ...
 
     @webmethod(route="/memory_bank/insert")
-    def insert_memory_documents(
+    def insert_documents(
         self,
         bank_id: str,
         documents: List[MemoryBankDocument],
     ) -> None: ...
 
     @webmethod(route="/memory_bank/update")
-    def update_memory_documents(
+    def update_documents(
         self,
         bank_id: str,
         documents: List[MemoryBankDocument],
     ) -> None: ...
 
-    @webmethod(route="/memory_bank/get")
-    def retrieve_memory_documents(
-        self,
-        request: RetrieveMemoryDocumentsRequest,
-    ) -> List[MemoryBankDocument]: ...
-
-    @webmethod(route="/memory_bank/get")
-    def get_memory_documents(
+    @webmethod(route="/memory_bank/query")
+    def query_documents(
         self,
         bank_id: str,
-        document_uuids: List[str],
-    ) -> List[MemoryBankDocument]: ...
+        query: InterleavedTextMedia,
+        params: Optional[Dict[str, Any]] = None,
+    ) -> QueryDocumentsResponse: ...
 
-    @webmethod(route="/memory_bank/delete")
-    def delete_memory_documents(
+    @webmethod(route="/memory_bank/documents/get")
+    def get_documents(
         self,
         bank_id: str,
-        document_uuids: List[str],
-    ) -> List[str]: ...
+        document_ids: List[str],
+    ) -> List[MemoryBankDocument]: ...
+
+    @webmethod(route="/memory_bank/documents/delete")
+    def delete_documents(
+        self,
+        bank_id: str,
+        document_ids: List[str],
+    ) -> None: ...