create_tagging_chain with examples or few shots

[miguelmedinaperez]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

I've modified the create_tagging_chain function to accept examples within the schema, as this enhancement significantly improves tagging accuracy. I propose incorporating this modification into LangChain. However, I'm open to alternative solutions if there's a way to achieve comparable results with the function's current implementation.

Motivation

In many cases, the accuracy of create_tagging_chain improves when including examples of the type input -> output; however, I have not found an easy way to do so with the current implementation of create_tagging_chain.

Proposal (If applicable)

This is my current proposal, which essentially allows for the direct inclusion of examples in the schema that will be passed to create_tagging_chain.

product_metadata_schema = {
    "properties": {
        "fashion_category": {
            "type": "string",
            "description": "Fashion category of the 'fashion_product'. Examples: 'trousers', 'shoes', 'dress', 'underwear', 'jewelry', 'accessories', 'uniform'. Infer a category not to general as 'fashion' or 'retail'.",
        },
        "fashion_product": {
            "type": "string",
            "description": "Minimal product description (usually a noun and seldom including an adjective). Examples: 'jeans', 'moccasins', 'mini skirt', 'necklace', 'onesie'. This field is mandatory.",
        },
        "brand": {
            "type": "string",
            "description": "Product brand. If you have doubts, do not include this field.",
        },
        "color": {
            "type": "string",
            "description": "Product color. If you have doubts, do not include this field.",
        },
        "gender": {
            "type": "string",
            "description": "Target gender of the product. If you have doubts, do not include this field.",
            "enum": ["male", "female", "unisex"],
        },
        "age_range": {
            "type": "string",
            "description": """
            Target age range of the product. If you have doubts, do not include this field. Follow the next rules:
            - If 'man', 'woman', 'sir' or 'madam', then 'adult'.
            - If 'boy' or 'girl' then 'youth'.
            - If 'child', 'school-age', or 'toddler' then 'child'.
            - If 'baby', 'newborn' or 'infant' then 'baby'.
            """,
            "enum": ["adult", "youth", "child", "baby"],
        },
    },
    "required": ["fashion_product"],
    "examples": [
        {
            "input": "Quilted vest with flower print in green and mustard for girls",
            "tags": {
                "fashion_category": "sweaters and jackets",
                "fashion_product": "vest",
                "color": "green and mustard",
                "gender": "female",
                "age_range": "youth",
            },
        },
        {
            "input": "Tommy men's belt",
            "tags": {
                "fashion_category": "accessories",
                "fashion_product": "belt",
                "brand": "Tommy",
                "gender": "male",
                "age_range": "adult",
            },
        },
        {
            "input": "Lotus 18106/3 Marc Márquez children's rubber watch",
            "tags": {
                "fashion_category": "watchmaking",
                "fashion_product": "watch",
                "brand": "Lotus",
                "gender": "male",
                "age_range": "child",
            },
        },
        {
            "input": "Black pantyhose with shaping girdle",
            "tags": {
                "fashion_category": "underwear",
                "fashion_product": "pantyhose",
                "color": "black",
                "gender": "female",
                "age_range": "adult",
            },
        },
        {
            "input": "Baby boy's blue jacket by Tizzas",
            "tags": {
                "fashion_category": "sweaters and jackets",
                "fashion_product": "jacket",
                "brand": "Tizzas",
                "color": "blue",
                "gender": "male",
                "age_range": "baby",
            },
        },
        {
            "input": "Men's multicolor undershirt",
            "tags": {
                "fashion_category": "underwear",
                "fashion_product": "undershirt",
                "color": "multicolor",
                "gender": "male",
                "age_range": "adult",
            },
        },
    ],
}


def _convert_schema(schema: dict) -> dict:
    props = {k: {"title": k, **v} for k, v in schema["properties"].items()}
    return {
        "type": "object",
        "properties": props,
        "required": schema.get("required", []),
        "examples": schema.get("examples", []),
    }


def _get_tagging_function(schema: dict) -> dict:
    return {
        "name": "information_extraction",
        "description": "Extracts the relevant information from the passage.",
        "parameters": _convert_schema(schema),
        "examples": _convert_schema(schema)["examples"],
    }


_TAGGING_TEMPLATE = """Extract the desired information from the following passage.

Only extract the properties mentioned in the 'information_extraction' function.

Examples:
{examples}

Passage:
{input}
"""


def format_examples(examples: list) -> str:
    formatted_examples = ""
    for example in examples:
        formatted_examples += f"- input: '{example['input']}'\n  output: ("
        for tag, value in example["tags"].items():
            formatted_examples += f"'{tag}': '{value}', "
        formatted_examples = formatted_examples[:-2] + ")\n"
    return formatted_examples


def create_tagging_chain(
    schema: dict,
    llm: BaseLanguageModel,
    prompt: Optional[ChatPromptTemplate] = None,
    **kwargs: Any,
) -> Chain:
    function = _get_tagging_function(schema)
    examples_formatted = format_examples(schema.get("examples", []))
    # Update the template to include the formatted examples
    custom_template = _TAGGING_TEMPLATE.format(
        examples=examples_formatted, input="{input}"
    )
    prompt = prompt or ChatPromptTemplate.from_template(custom_template)
    output_parser = JsonOutputFunctionsParser()
    llm_kwargs = get_llm_kwargs(function)
    chain = LLMChain(
        llm=llm,
        prompt=prompt,
        llm_kwargs=llm_kwargs,
        output_parser=output_parser,
        **kwargs,
    )
    return chain


chain = create_tagging_chain(
    schema=product_metadata_schema,
    llm=azure_gpt35,
)

descriptions = [
    "Baby boy swimsuit with stars",
    "Boy's Star Wars T-shirt 100% organic cotton",
    "Boy's skater print T-shirt",
    "Men's Easy Iron regular fit plain white shirt",
    "Basic girl's sweatshirt with print",
    "Baby girl's T-shirt with appliques",
    "Boy's ikat style boxer swimsuit"
]

for d in descriptions:
    try:
        product_metadata = chain.run(d)
        print(d)
        print(product_metadata)
    except Exception as e:
        print(f"Error processing '{d}': {e}")
    print("*******************************")