Serapeum Core

Serapeum Core is the provider-agnostic foundation for the Serapeum ecosystem. It defines the core LLM abstractions, prompt templates, structured-output parsers, and tool-execution utilities used by integration packages (OpenAI, Ollama, etc.).

If you are implementing a new LLM backend or want a consistent set of models and helpers for prompts, output parsing, and tool calling, this is the package to use.

Highlights

• Unified LLM interfaces for chat/completions, streaming, and async variants.
• Prompt templates for string and chat workflows with variable mapping utilities.
• Structured output parsing with Pydantic models and JSON schema hints.
• Tool metadata, schema generation, and safe tool execution utilities.
• Orchestrators for tool calling and structured outputs.

Package layout

• serapeum.core.base.llms: Base interfaces and data models (messages, chunks, responses, metadata).
• serapeum.core.llms: High-level LLM class with predict/stream helpers and structured output utilities.
• serapeum.core.prompts: Prompt templates for text and chat models.
• serapeum.core.output_parsers: Parsers for structured outputs (e.g., Pydantic).
• serapeum.core.tools: Tool metadata, tool interfaces, and tool execution.
• serapeum.core.llms.orchestrators: Higher-level programs for structured outputs and function-calling orchestration.
• serapeum.core.utils: JSON/schema helpers and async utilities.
• serapeum.core.configs: Global configuration singleton (Configs).
• serapeum.core.chat: Agent response container with streaming helpers.

Installation

From the repo root:

cd libs/core
python -m pip install -e .

Python 3.11+ is required.

Quick start

1) Build a minimal LLM implementation

from serapeum.core.llms import LLM, CompletionResponse, Metadata
from serapeum.core.prompts import PromptTemplate


class EchoLLM(LLM):
  metadata = Metadata.model_construct(is_chat_model=False)

  def chat(self, messages, stream=False, **kwargs):
    raise NotImplementedError()

  async def achat(self, messages, stream=False, **kwargs):
    raise NotImplementedError()

  def complete(self, prompt, formatted=False, stream=False, **kwargs):
    if stream:
      raise NotImplementedError()
    return CompletionResponse(text=prompt, delta=prompt)

  async def acomplete(self, prompt, formatted=False, stream=False, **kwargs):
    if stream:
      raise NotImplementedError()
    return CompletionResponse(text=prompt, delta=prompt)


llm = EchoLLM()
prompt = PromptTemplate("Hello, {name}!")
result = llm.predict(prompt, name="Serapeum")
print(result)

2) Parse structured outputs with Pydantic

from pydantic import BaseModel
from serapeum.core.output_parsers import PydanticParser
from serapeum.core.prompts import PromptTemplate


class Greeting(BaseModel):
    message: str


parser = PydanticParser(output_cls=Greeting)
prompt = PromptTemplate(
    'Return JSON like {"message": "<text>"}. Text: {text}',
    output_parser=parser,
)

# With a real LLM backend, this returns a validated Greeting model.
# result = llm.predict(prompt, text="Hello")

3) Define tools and execute them safely

from serapeum.core.tools import BaseTool, ToolMetadata, ToolOutput, ToolExecutor


class EchoTool(BaseTool):
    @property
    def metadata(self) -> ToolMetadata:
        return ToolMetadata(description="Echo input", name="echo")

    def __call__(self, input_values: dict) -> ToolOutput:
        return ToolOutput(tool_name="echo", content=input_values.get("input", ""))


tool = EchoTool()
executor = ToolExecutor()
output = executor.execute(tool, {"input": "hi"})
print(output.content)

Core concepts

LLM interface and orchestration

• BaseLLM defines the provider contract for chat/completion endpoints, streaming, and async variants.
• LLM builds on BaseLLM with helpers for prompt formatting, prediction, and structured output modes.
• FunctionCallingLLM extends LLM with tool-calling helper methods.

Prompts

Use PromptTemplate for string prompts and ChatPromptTemplate for message templates. Both support template variable mappings and optional output parsers.

Output parsers

PydanticParser injects a compact JSON schema into prompts and parses LLM output into validated Pydantic models. output_parsers.utils also provides markdown JSON/code extraction helpers.

Tools and schemas

• ToolMetadata produces OpenAI-style tool specs and JSON schema for tool inputs.
• CallableTool (in serapeum.core.tools.callable_tool) can wrap functions or Pydantic models into tool definitions.
• ToolExecutor runs tools safely with optional auto-unpacking and error normalization.

Structured programs

• TextCompletionLLM runs a prompt + parser + LLM pipeline to return Pydantic outputs for completion-style models.
• ToolOrchestratingLLM uses function-calling models to produce structured outputs via tools.

Configuration

serapeum.core.configs.Configs is a small global configuration holder. You can set a default LLM instance and control structured output mode:

from serapeum.core.configs import Configs
from serapeum.core.types import StructuredOutputMode


Configs.llm = llm
Configs.structured_output_mode = StructuredOutputMode.OPENAI

Links

• Homepage: https://github.com/Serapieum-of-alex/serapeum
• Docs: https://serapeum.readthedocs.io/
• Changelog: https://github.com/Serapieum-of-alex/serapeum/HISTORY.rst