The following objects in Rigging have great serialization support for storage and retrieval:

  • Chat
  • Completion
  • Generator
  • Model

Most of this stems from our use of Pydantic for core models, and we’ve included some helpful fields for reconstructing Chats and Completions.

​
JSON Serialization

Let’s build a joke pipeline and serialize the final chat into JSON.

import rigging as rg

class Joke(rg.Model):
    content: str

chat = (
    await
    rg.get_generator("gpt-3.5-turbo")
    .chat(f"Provide 3 jokes each between {Joke.xml_tags()} tags.")
    .meta(tags=['joke'])
    .with_(temperature=1.25)
    .run()
)

chat.last.parse_set(Joke)

serialized = chat.model_dump_json(indent=2)
print(serialized)

You’ll notice that every Chat gets a unique id field to help track them in a datastore like Elastic or Pandas. We also assign a timestamp to understand when the generation took place. We are also taking advantage of the .meta() rigging.chat.ChatPipeline.meta to add a tracking tag for filtering later.

​
JSON Deserialization

The JSON has everything required to reconstruct a Chat including a generator_id dynamically constructed to preserve the parameters used to create the generated message(s). We can now deserialize a chat from a datastore, and immediately step back into a ChatPipeline for exploration.

chat = rg.Chat.model_validate_json(serialized)
print(chat.conversation)
# [user]: Provide 3 jokes each between <joke></joke> tags.

# [assistant]:
# <joke> Why was the math book sad? Because it had too many problems. </joke>
# <joke> I told my wife she should embrace her mistakes. She gave me a hug. </joke>
# <joke> Why did the scarecrow win an award? Because he was outstanding in his field. </joke>

continued = chat.continue_("Please explain the first joke to me.").run()
print(continued.last)
# [assistant]: In the first joke, the pun is based on the double meaning of the word "problems."
# The math book is described as being sad because it has "too many problems," which could be
# interpreted as having both mathematical problems (equations to solve) and emotional difficulties.
# This play on words adds humor to the joke.

​
Pandas DataFrames

Rigging also has helpers in the rigging.data module for performing conversions between Chat objects and other storage formats like Pandas. In chats_to_df the messages are flattened and stored with a chat_id column for grouping. df_to_chats allows you to reconstruct a list of Chat objects back from a DataFrame.

import rigging as rg

chats = (
    await
    rg.get_generator("claude-3-haiku-20240307")
    .chat("Write me a haiku.")
    .run_many(3)
)

df = rg.data.chats_to_df(chats)
# or
df = chats.to_df()

print(df.info())

# RangeIndex: 6 entries, 0 to 5
# Data columns (total 9 columns):
#  #   Column             Non-Null Count  Dtype
# ---  ------             --------------  -----
#  0   chat_id            6 non-null      string
#  1   chat_metadata      6 non-null      string
#  2   chat_generator_id  6 non-null      string
#  3   chat_timestamp     6 non-null      datetime64[ms]
#  4   generated          6 non-null      bool
#  5   role               6 non-null      category
#  6   parts              6 non-null      string
#  7   content            6 non-null      string
#  8   message_id         6 non-null      string
# dtypes: bool(1), category(1), datetime64[ms](1), string(6)

df.content.apply(lambda x: len(x)).mean()

# 60.166666666666664

back = rg.data.df_to_chats(df)
print(back[0].conversation)

# [user]: Write me a haiku.
#
# [assistant]: Here's a haiku for you:
#
# Gentle breeze whispers,
# Flowers bloom in vibrant hues,
# Nature's simple bliss.