Rigging has good support for iterating over messages, parameters, and generators, as well as large batching of requests. How efficiently these mechanisms operates is dependent on the underlying generator that’s being used, but Rigging has been developed with scale in mind.

​
Multiple Generations

The run_many functions let you scale out generation N times with the same inputs:

  • ChatPipeline.run_many()
  • CompletionPipeline.run_many()
  • Prompt.run_many()
import rigging as rg

async def check_animal(chats: list[rg.Chat]) -> list[rg.Chat]:
    return [
        await chat.continue_(f"Why did you pick that animal?").meta(questioned=True).run()
        if any(a in chat.last.content.lower() for a in ["cat", "dog", "cow", "mouse"])
        else chat
        for chat in chats
    ]

chats = (
    await
    rg.get_generator("gpt-3.5-turbo")
    .chat("Tell me a joke about an animal.")
    .map(check_animal)
    .run_many(3)
)

for i, chat in enumerate(chats):
    questioned = chat.metadata.get("questioned", False)
    print(f"--- Chat {i+1} (?: {questioned}) ---")
    print(chat.conversation)
    print()

​
Batching Inputs

The run_batch functions let you batch across a set of inputs:

  • ChatPipeline.run_batch()
  • CompletionPipeline.run_batch()

As processing proceeds with things like .then() or .map(), the chats will resolve individually and collapse into the final results.

import rigging as rg
from rigging.model import CommaDelimitedAnswer

pipeline = (
    rg.get_generator('gpt-4-turbo')
    .chat({
        "role": "system",
        "content": f"Always respond with {CommaDelimitedAnswer.xml_tags()} tags."}
    )
    .until_parsed_as(CommaDelimitedAnswer, attempt_recovery=True)
)

many = [f"Give me 3 famous {thing}" for thing in ["authors", "painters", "musicians", "hackers"]]

chats = await pipeline.run_batch(many, on_failed='skip')

for i, chat in enumerate(chats):
    print(f"--- Chat {i+1} ({len(chat)}) ---")
    print(chat.last.parse(CommaDelimitedAnswer).items)
    print()

Skipping failed results

Passing on_failed='skip' to .run_batch, or configuring a pipeline with .catch(..., on_failed='skip') will cause the function to ignore any parsing errors like ExhaustedMaxRoundsError and only return successful chats.

​
Batching Parameters

In addition to batching against input messages or strings, you can fix a single input and build a batch across a set of generation parameters. The inputs to .run_batch will scale either the generate parameters or the input messages if either is a single item.

import rigging as rg

pipeline = rg.get_generator("gpt-3.5-turbo").chat()

chats = await pipeline.run_batch(
    ["Tell me a short fact about an japanese city."],
    [rg.GenerateParams(temperature=t) for t in [0.6, 0.9, 1.2, 1.5, 1.8]]
)

for i, chat in enumerate(chats):
    print(f"--- Chat {i+1} ---")
    print(chat.generator_id)
    print()
    print(chat.conversation)
    print()

​
Iterating over Models

The run_over functions let you execute generation over a set of generators:

  • ChatPipeline.run_over()
  • CompletionPipeline.run_over()
  • Prompt.run_over()

Generators can be passed as string identifiers or full instances of Generator. By default the original generator associated with the ChatPipeline is included in the iteration, configurable with the include_original parameter.

Much like the run_many and run_batch functions, you can control the handling of failures with the on_failed parameter.

import rigging as rg
from rigging.model import Answer

QUESTION = "What is the capital of France?"
ANSWER = "paris"

async def score_output(chats: list[rg.Chat]) -> list[rg.Chat]:
    return [
        chat.meta(correct=chat.last.parse(Answer).content.lower() == ANSWER)
        for chat in chats
    ]

chats = (
    await
    rg.get_generator("gpt-3.5-turbo")
    .chat([
        {"role": "system", "content": f"Always respond in one word between {Answer.xml_tags()} tags."},
        {"role": "user", "content": QUESTION}
    ])
    .until_parsed_as(Answer, max_rounds=3)
    .map(score_output)
    .run_over("gpt-4-turbo", "claude-3-haiku-20240307,temperature=0.5", "claude-3-sonnet-20240229")
)

for chat in chats:
    print("Model: ", chat.generator.model)
    print("Msg:   ", chat.last.content)
    print("Meta:  ", chat.metadata)
    print()