Implementation:Ucbepic Docetl Directive Registry

Knowledge Sources	Ucbepic_Docetl
Domains	Pipeline_Optimization, LLM_Operations
Last Updated	2026-02-08 00:00 GMT

Overview

Concrete registry module for all available reasoning optimizer directives provided by the DocETL reasoning optimizer.

Description

The __init__.py module serves as the central registry for all directive classes in the reasoning optimizer. It imports all directive implementations, maintains several registries (ALL_DIRECTIVES, ALL_COST_DIRECTIVES, DIRECTIVE_GROUPS, MULTI_INSTANCE_DIRECTIVES), and provides a DIRECTIVE_REGISTRY name-to-instance mapping. It also exposes utility functions like get_all_directive_strings(), get_all_cost_directive_strings(), and instantiate_directive() for centralized directive lookup and instantiation.

Usage

This module is used by the MOAR agent and optimization pipeline to discover available directives, generate prompt descriptions of all directives, and instantiate directives by name. It is the single entry point for the directive subsystem.

Code Reference

Source Location

Repository: Ucbepic_Docetl
File: docetl/reasoning_optimizer/directives/__init__.py
Lines: 1-192

Signature

ALL_DIRECTIVES = [
    ChainingDirective(), GleaningDirective(), ReduceGleaningDirective(),
    ReduceChainingDirective(), ChangeModelCostDirective(), DocSummarizationDirective(),
    IsolatingSubtasksDirective(), DeterministicDocCompressionDirective(),
    OperatorFusionDirective(), DocumentChunkingDirective(), DocumentChunkingTopKDirective(),
    ChunkHeaderSummaryDirective(), TakeHeadTailDirective(), ClarifyInstructionsDirective(),
    SwapWithCodeDirective(), MapReduceFusionDirective(), HierarchicalReduceDirective(),
    CascadeFilteringDirective(), ArbitraryRewriteDirective(),
    MapToMapResolveReduceDirective(), MapResolveToMapWithCategoriesDirective(),
]

DIRECTIVE_REGISTRY = {directive.name: directive for directive in ALL_DIRECTIVES}

def get_all_directive_strings() -> str: ...
def get_all_cost_directive_strings() -> str: ...
def instantiate_directive(directive_name: str, operators: List[Dict], target_ops: List[str],
                          agent_llm: str, message_history: list, **kwargs): ...

Import

from docetl.reasoning_optimizer.directives import ALL_DIRECTIVES, DIRECTIVE_REGISTRY, instantiate_directive

I/O Contract

Inputs

Name	Type	Required	Description
directive_name	str	Yes	Name of the directive to instantiate
operators	List[Dict]	Yes	List of pipeline operators
target_ops	List[str]	Yes	List of target operation names
agent_llm	str	Yes	LLM model to use for instantiation
message_history	list	Yes	Conversation history

Outputs

Name	Type	Description
new_ops_list	list	Transformed operation configurations
updated_message_history	list	Updated conversation history after instantiation

Usage Examples

# Using the registry to instantiate a directive by name
from docetl.reasoning_optimizer.directives import instantiate_directive

new_ops, history = instantiate_directive(
    directive_name="chaining",
    operators=pipeline_operators,
    target_ops=["extract_contract_terms"],
    agent_llm="gpt-4.1",
    message_history=[]
)

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment