Assistant Service Kit

Reusable Python package for building LangChain-based assistant services with FastAPI, auth, tenancy, tools, RAG, tracing, storage, and observability.

Install

Base runtime:

pip install assistant-service-kit

Provider-specific installs:

pip install "assistant-service-kit[rag,gemini]"
pip install "assistant-service-kit[rag,openai]"
pip install "assistant-service-kit[rag,anthropic]"

For local development from this package directory:

pip install -e ".[dev]"

Core Pieces

• Assistant: simple facade for building a FastAPI assistant quickly.
• AssistantDefinition: declares the assistant server contract.
• AssistantEndpoints / EndpointPolicy: declares enabled routes, custom paths, names, global endpoint policy, and endpoint permissions.
• AuthConnector: supports no-auth, user-only, trusted internal token, API key, and custom identity resolution.
• LLMProviderRegistry: Gemini, OpenAI, Anthropic, and fake chat model providers.
• EmbeddingProviderRegistry: Gemini, OpenAI, Anthropic, and fake embedding providers.
• ApiClient / ApiClientConfig: reusable HTTP client for tools, including API key, bearer, forwarded token, and gateway-delegated token auth.
• ToolRegistry / ToolPolicy: exposes and validates tools by tenant, user, roles, and limits.
• SQLAlchemyTraceStore / InMemoryTraceStore: reusable chat trace persistence.
• PostgresRagStore / VectorStoreConnector: reusable RAG storage and vector-store integration points.
• create_assistant_sqlalchemy_models: standard SQLAlchemy trace and RAG models for the common case.
• answer_with_langchain_agent: LangChain tool-calling orchestrator with observability metadata.

Minimal Example

from assistant_service_kit import (
    Assistant,
    AssistantEndpoints,
    DatabaseConfig,
    TenancyMode,
    create_sqlalchemy_database,
)

database = create_sqlalchemy_database(DatabaseConfig(debug=True, debug_url="sqlite:///:memory:"))

assistant = Assistant(
    name="demo",
    title="Demo Assistant",
    base_path="/api/assistant",
    tenancy=TenancyMode.NONE,
    prompt="Answer clearly.",
    endpoints=AssistantEndpoints(chat=False, conversations=False, messages=False, feedback=False),
)

assistant.with_no_auth()
assistant.with_llm(provider="fake", model="fake", extra={"response": "ok"})
assistant.with_database(database)

app = assistant.fastapi()

Declarative tools:

from pydantic import BaseModel
from assistant_service_kit import ApiClientConfig, ToolPolicy, ToolRegistry


class SearchArgs(BaseModel):
    query: str


tools = ToolRegistry()
tools.api_tool(
    name="search_data",
    description="Search tenant data.",
    args_schema=SearchArgs,
    policy=ToolPolicy(requires_tenant=True, requires_user=True, max_limit=10),
    client="data_api",
    method="GET",
    path="/internal/search/",
    params="payload",
)


assistant.with_api_client("data_api", ApiClientConfig(base_url="https://internal-api.example"))
assistant.with_tools(tools)

Use @tools.tool(...) when a tool needs custom Python logic instead of a direct API call.

For standard SQLAlchemy persistence:

from assistant_service_kit import DatabaseConfig, create_sqlalchemy_database

database = create_sqlalchemy_database(
    DatabaseConfig(
        debug=True,
        debug_url="sqlite:///./assistant.db",
        url="postgresql+psycopg2://user:pass@host/db",
    )
)

assistant.with_database(database)
assistant.with_rag(database, default_limit=4, max_limit=8, chunk_size=800, chunk_overlap=120)
app = assistant.fastapi()

For advanced integrations, call assistant.build() to get an AssistantDefinition, or build AssistantDefinition and connectors directly.

Status

The package is in alpha. APIs can change until the first stable release.