GrammarFlow 0.0.5

Ensuring parsability of LLM responses in agent chains

🪢 GrammarFlow

🚀 Supercharging agent chains with constrained LLM outputs 🚀

🤔 What is this?

This repository contains code to abstract the LLM output constraining process. It helps you define your grammar rules using Pydantic and Typing in a pythonic way, and inherently embeds metadata from these dataclasses into the prompt. Parsing is enabled in JSON, TOML and XML formats, with custom parsers that avoid the issues faced by json.loads (..etc) while parsing direct outputs. It can also create GNBF grammr from the same, which is used by the llama.cpp package for sampling logits smartly.

The goal of this package was to overcome the issues faced when using langchain's output parsers with local language models. While GPT-4 produces consistent results in returning the correct formats, Llama-7B would cause parsing errors in my testing chains with more complex prompts.

📚 Quick Install

pip install grammarflow

📃 Code Usage

1. Map out what your agent chain is doing. Understand what it's goals are and what data needs to be carried forward from one step to the next. For example, consider the ReAct prompting framework. In every call, we want to pass in the Action and subsequent Observation to the next call.

2. Make a Pydantic Model for the above case. Here's a sample:

class ThoughtState(BaseModel):
    thought: str
    goal: str
    tool: str = Field(...,
                      description="Choose one of ['Web_QA', 'Web_Search', 'Web_Scraping', 'Web_Automation', 'Web_Research']")
    action: str = Field(...,
                        description="Choose one of ['Create', 'Update', 'Delete', 'Read']")
    action_input: str = Field(..., description="The input data for the action")
    thought_id: Optional[str] = Field(
        None, description="1 if it is the first thought, 0 if it is the final thought.")

3. [Optional] Create a prompt template using grammarflow's PromptBuilder. Below is an example of the Llama prompt template.

from grammarflow import PromptBuilder
llama_prompt = PromptBuilder()
llama_prompt.add_section(
    text="<s>[INST] <<SYS>>\n{system_context}\n<</SYS>>",
    placeholder="system_context",
    define_grammar=True,
)
llama_prompt.add_section(
    text="{user_message}[/INST]",
    placeholder="user_message",
)

You can find an in-depth explanation on making prompts here!

4. [Optional] If you decide to make your own template, define your system_context and user_message placeholders.

system_context = """Your goal is to think and plan out how to solve questions using agent tools provided to you. Think about all aspects of your thought process."""
user_message = """Who is Vladmir Putin?"""

5. Invoke the Constrain block with the prompt. Set the configuration metadata, and format the prompt with the required grammars and placeholders.

from grammarflow import Constrain 

with Constrain(llama_prompt) as manager:
    manager.set_config(
        format='json', # or 'xml', 'toml'. 
        return_sequence='single_response' # or 'multi_response', if you need multiple grammars. 
    )

    # Makes the changes to the prompt
    manager.format_prompt(placeholders={ # if you have placeholders in the prompt
                          'user_message': user_message,
                          'system_context': system_context
                          },
                          grammars=[{
                              'description': 'This format describes your current thinking state', # Description of the response format
                              'model': [ThoughtState]}
                          ]
    )

    # Assume `llm` to be a call to a model
    llm_response = llm.request(manager.prompt, temperature=0.01)

    # Parse the response into a custom dataclass for holding values
    response = manager.parse(llm_response)

6. Extract the required values from the response to perform necessary functions on.

observation = PerformSomeAction(
  action = response.ThoughtState.action, 
  action_input = response.ThoughtState.action_input
) 

7. Continue to the next iteration in your agent chain!

GNBF Grammar

GrammarFlow also has functionality to convert a pydantic model into GNBF grammar.

"GBNF (GGML BNF) is a format for defining formal grammars to constrain model outputs in llama.cpp. For example, you can use it to force the model to generate valid JSON, or speak only in emojis." Read more about it here: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md

This can then be passed into llama.cpp using llama-cpp-python to ensure that sampling of logits can take place with rules in mind.

# Define your model 
class TeamMember(BaseModel):
    name: str
    role: str

class TaskUpdate(BaseModel):
    update_time: float
    comment: Optional[str] = None
    status: bool

class Task(BaseModel):
    title: str
    description: str
    assigned_to: List[TeamMember]
    due_date: List[str]
    updates: List[TaskUpdate]

class Project(BaseModel):
    name: str
    description: str
    project_url: Optional[str] = None
    team_members: List[TeamMember]
    grammars: Task

# Convert to grammar
from grammarflow import GNBF

grammar = GNBF(Project).generate_grammar()

# Verify with LlamaGrammar
GNBF.verify_grammar(grammar)

# Use it with the model 
with Constrain(llama_prompt) as manager: 
    manager.set_config(...)
    manager.format_prompt(...)

    llm_response = llm(
        manager.prompt,
        grammar=grammar, max_tokens=-1
    )
    response = manager.parse(llm_response)

Citation

We appreciate it if you would please cite this repo if you found the library useful for your work:

@software{GrammarFlow,
  author = {Ravikiran, Akshath Raghav and Culurciello, Eugenio},
  title = {GrammarFlow: Powering Agent Chains by Constraining LLM Outputs},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/e-lab/GrammarFlow}}, 
  version = {0.0.1}
}