fabricatio-core 0.8.10

A foundational Python library providing core components for building LLM-driven applications using an event-based agent structure.

fabricatio-core

MIT

Foundational Python library for building LLM-driven applications using an event-based agent architecture. Built on a hybrid Rust/Python foundation for performance-critical operations.

Requires Python 3.12+.

Installation

pip install fabricatio
# or
uv pip install fabricatio

For all optional extras:

pip install fabricatio[full]

Key Components

Event System (Event, EventEmitter)

Segmented event type with task-status lifecycle methods (push_running, push_finished, etc.) and glob-wildcard (*) matching. The global EventEmitter dispatches events asynchronously with on/off/emit.

from fabricatio_core.emitter import EMITTER
from fabricatio_core.rust import Event

EMITTER.on("task::*::finished", my_handler)

LLM Routing (Router, RouterUsage)

Multi-provider router for completion, embedding, and reranking. RouterUsage provides structured LLM interaction patterns — ask, list strings, code generation, judging, choosing — with configurable validation and fallback defaults.

from fabricatio_core.rust import ROUTER, router_usage

response = await ROUTER.completion(send_to="default", message="Hello")
result = await router_usage.choosing("Pick two items", valid_names=["a", "b", "c"], k=2)

Task Model (Task)

Pydantic-based task with lifecycle status management (Pending, Running, Finished, Failed, Cancelled), dependency tracking, structured proposal generation from prompts, and automatic event emission on state transitions.

task = Task(name="data_ingestion", description="Ingest raw data")
task.start()
# ... execute ...
task.finish(output={"rows": 1042})

Workflow Engine (Action, WorkFlow)

Abstract action pipeline with shared context propagation. WorkFlow orchestrates sequences of Action instances, managing input/output keys, error handling, and task lifecycle.

class ParseData(Action):
    async def _execute(self, **cxt):
        return parse(cxt["task_input"])

wf = WorkFlow(actions=[ParseData().to_task_output()], name="parse_pipeline")
await wf.execute(task)

Role Framework (Role)

Agent roles map event patterns to workflows. Global registry supports register_role / get_registered_role. Roles dispatch workflows automatically when matching events are emitted.

from fabricatio_core.models.role import Role, register_role

analyst = Role(name="analyst", description="Data analysis agent")
analyst.deploy_on("data::ingested", parse_wf)
register_role(analyst)

Template Rendering (TemplateManager)

Handlebars-based template engine. Load templates from directories, render with structured data.

from fabricatio_core.rust import TEMPLATE_MANAGER

TEMPLATE_MANAGER.add_store("./templates").discover_templates()
output = TEMPLATE_MANAGER.render_template("greeting", {"name": "World"})

Capability Mixins (UseLLM, UseEmbedding, UseReranker, Propose)

Inheritable classes that add LLM querying, embedding generation, reranking, and structured proposal capabilities to any model. All integrate with the Router for provider dispatch.

class MyAgent(UseLLM, UseEmbedding, Propose, WithBriefing):
    pass

agent = MyAgent(name="helper")
answer = await agent.aask(question="What is 2+2?")
obj = await agent.propose(MyModel, prompt="Create a config object")

Text Parsing & Language Utilities (Rust)

Fast Rust-backed functions for text processing:

• split_sentence_bounds / split_word_bounds — Unicode-aware text splitting
• split_into_chunks — chunk text with configurable overlap
• tokens_of / word_count — token and word counting
• blake3_hash — BLAKE3 content hashing
• detect_language — language detection
• is_english, is_chinese, is_japanese, etc. — language checks
• is_likely_text — file content type detection
• CodeSnippetParser, CodeBlockParser, GenericBlockParser, JsonParser — structured block extraction from LLM outputs

Base Model Hierarchy (models.generic)

Pydantic-based abstract base classes for consistent model design: Named, Described, WithBriefing, WithDependency, ScopedConfig (with hierarchical fallback), ProposedAble, InstantiateFromString, Display, and more.

Decorators

• cfg_on / cfg_on_async — feature-gated function execution
• depend_on_external_cmd — check external binary availability
• logging_execution_info / logging_exec_time — execution logging
• once — ensure single invocation

Configuration (Config)

Structured configuration with sub-sections for LLM, embedding, reranking, emitter, routing, templates, debug, and deployment settings. Accessible via fabricatio_core.rust.CONFIG.

Package Structure

fabricatio-core/
├── python/fabricatio_core/
│   ├── capabilities/      - LLM, embedding, reranker, and proposal mixins
│   ├── models/            - Action, Role, Task, generic base models, kwargs types
│   ├── rust/              - PyO3 Rust extension (autogenerated stubs)
│   ├── decorators.py      - cfg_on, once, logging_exec_time, etc.
│   ├── emitter.py         - EventEmitter with wildcard pattern matching
│   ├── journal.py         - Logger bridge to Rust logger
│   ├── utils.py           - ok, cfg, override_kwargs, first_available, etc.
│   └── __init__.py        - Public API surface
├── src/                   - Rust sources (parser, templates, event, language, etc.)
└── pyproject.toml

License

MIT — see LICENSE