Skip to content

srush/MiniChain

Repository files navigation

A tiny library for coding with large language models. Check out the MiniChain Zoo to get a sense of how it works.

Coding

  • Code (math_demo.py): Annotate Python functions that call language models.
@prompt(OpenAI(), template_file="math.pmpt.tpl")
def math_prompt(model, question):
    "Prompt to call GPT with a Jinja template"
    return model(dict(question=question))

@prompt(Python(), template="import math\n{{code}}")
def python(model, code):
    "Prompt to call Python interpreter"
    code = "\n".join(code.strip().split("\n")[1:-1])
    return model(dict(code=code))

def math_demo(question):
    "Chain them together"
    return python(math_prompt(question))
  • Chains (Space): MiniChain builds a graph (think like PyTorch) of all the calls you make for debugging and error handling.

show(math_demo,
     examples=["What is the sum of the powers of 3 (3^i) that are smaller than 100?",
               "What is the sum of the 10 first positive integers?"],
     subprompts=[math_prompt, python],
     out_type="markdown").queue().launch()
...
Question:
A robe takes 2 bolts of blue fiber and half that much white fiber. How many bolts in total does it take?
Code:
2 + 2/2

Question:
{{question}}
Code:
  • Installation
pip install minichain
export OPENAI_API_KEY="sk-***"

Examples

This library allows us to implement several popular approaches in a few lines of code.

It supports the current backends.

  • OpenAI (Completions / Embeddings)
  • Hugging Face πŸ€—
  • Google Search
  • Python
  • Manifest-ML (AI21, Cohere, Together)
  • Bash

Why Mini-Chain?

There are several very popular libraries for prompt chaining, notably: LangChain, Promptify, and GPTIndex. These library are useful, but they are extremely large and complex. MiniChain aims to implement the core prompt chaining functionality in a tiny digestable library.

Tutorial

Mini-chain is based on annotating functions as prompts.

image

@prompt(OpenAI())
def color_prompt(model, input):
    return model(f"Answer 'Yes' if this is a color, {input}. Answer:")

Prompt functions act like python functions, except they are lazy to access the result you need to call run().

if color_prompt("blue").run() == "Yes":
    print("It's a color")

Alternatively you can chain prompts together. Prompts are lazy, so if you want to manipulate them you need to add @transform() to your function. For example:

@transform()
def said_yes(input):
    return input == "Yes"

image

@prompt(OpenAI())
def adjective_prompt(model, input):
    return model(f"Give an adjective to describe {input}. Answer:")
adjective = adjective_prompt("rainbow")
if said_yes(color_prompt(adjective)).run():
    print("It's a color")

We also include an argument template_file which assumes model uses template from the Jinja language. This allows us to separate prompt text from the python code.

@prompt(OpenAI(), template_file="math.pmpt.tpl")
def math_prompt(model, question):
    return model(dict(question=question))

Visualization

MiniChain has a built-in prompt visualization system using Gradio. If you construct a function that calls a prompt chain you can visualize it by calling show and launch. This can be done directly in a notebook as well.

show(math_demo,
     examples=["What is the sum of the powers of 3 (3^i) that are smaller than 100?",
              "What is the sum of the 10 first positive integers?"],
     subprompts=[math_prompt, python],
     out_type="markdown").queue().launch()

Memory

MiniChain does not build in an explicit stateful memory class. We recommend implementing it as a queue.

image

Here is a class you might find useful to keep track of responses.

@dataclass
class State:
    memory: List[Tuple[str, str]]
    human_input: str = ""

    def push(self, response: str) -> "State":
        memory = self.memory if len(self.memory) < MEMORY_LIMIT else self.memory[1:]
        return State(memory + [(self.human_input, response)])

See the full Chat example. It keeps track of the last two responses that it has seen.

Tools and agents.

MiniChain does not provide agents or tools. If you want that functionality you can use the tool_num argument of model which allows you to select from multiple different possible backends. It's easy to add new backends of your own (see the GradioExample).

@prompt([Python(), Bash()])
def math_prompt(model, input, lang):
    return model(input, tool_num= 0 if lang == "python" else 1)

Documents and Embeddings

MiniChain does not manage documents and embeddings. We recommend using the Hugging Face Datasets library with built in FAISS indexing.

image

Here is the implementation.

# Load and index a dataset
olympics = datasets.load_from_disk("olympics.data")
olympics.add_faiss_index("embeddings")

@prompt(OpenAIEmbed())
def get_neighbors(model, inp, k):
    embedding = model(inp)
    res = olympics.get_nearest_examples("embeddings", np.array(embedding), k)
    return res.examples["content"]

This creates a K-nearest neighbors (KNN) prompt that looks up the 3 closest documents based on embeddings of the question asked. See the full Retrieval-Augemented QA example.

We recommend creating these embeddings offline using the batch map functionality of the datasets library.

def embed(x):
    emb = openai.Embedding.create(input=x["content"], engine=EMBEDDING_MODEL)
    return {"embeddings": [np.array(emb['data'][i]['embedding'])
                           for i in range(len(emb["data"]))]}
x = dataset.map(embed, batch_size=BATCH_SIZE, batched=True)
x.save_to_disk("olympics.data")

There are other ways to do this such as sqllite or Weaviate.

Typed Prompts

MiniChain can automatically generate a prompt header for you that aims to ensure the output follows a given typed specification. For example, if you run the following code MiniChain will produce prompt that returns a list of Player objects.

class StatType(Enum):
    POINTS = 1
    REBOUNDS = 2
    ASSISTS = 3

@dataclass
class Stat:
    value: int
    stat: StatType

@dataclass
class Player:
    player: str
    stats: List[Stat]


@prompt(OpenAI(), template_file="stats.pmpt.tpl", parser="json")
def stats(model, passage):
    out = model(dict(passage=passage, typ=type_to_prompt(Player)))
    return [Player(**j) for j in out]

Specifically it will provide your template with a string typ that you can use. For this example the string will be of the following form:

You are a highly intelligent and accurate information extraction system. You take passage as input and your task is to find parts of the passage to answer questions.

You need to output a list of JSON encoded values

You need to classify in to the following types for key: "color":

RED
GREEN
BLUE


Only select from the above list, or "Other".⏎


You need to classify in to the following types for key: "object":⏎

String



You need to classify in to the following types for key: "explanation":

String

[{ "color" : "color" ,  "object" : "object" ,  "explanation" : "explanation"}, ...]

Make sure every output is exactly seen in the document. Find as many as you can.

This will then be converted to an object automatically for you.