Evaluating multimodal receipt extraction

Sep 30, 2024

Document extraction is a use case that is near and dear to my heart. The last time I dug deeply into it, there were not nearly as many models capable of solving for it as there are today. In honor of Pixtral and LLaMa3.2, I thought it would be fun to revisit it with the classic SROIE dataset.

The results are fascinating:

GPT-4o-mini performs the best, even better than GPT-4o
Pixtral 12B is almost as good as LLaMa 3.2 90B
The LLaMa models are almost 3x faster than the alternatives

Scatter plot

Let's jump right in!

Install dependencies

%pip install autoevals braintrust requests openai

Setup LLM clients

We'll use OpenAI's GPT-4o, LLaMa 3.2 11B and 90B, and Pixtral 12B with a bunch of test cases from SROIE and see how they perform. You can access each of these models behind the vanilla OpenAI client using Braintrust's proxy.

import os
 
import braintrust
import openai
 
client = braintrust.wrap_openai(
    openai.AsyncOpenAI(
        api_key=os.environ["BRAINTRUST_API_KEY"],
        base_url="https://api.braintrust.dev/v1/proxy",
    )
)

Downloading the data and sanity testing it

The zzzDavid/ICDAR-2019-SROIE repo has neatly organized the data for us. The files are enumerated in a 3 digit convention and for each image (e.g. 002.jpg), there is a corresponding file (e.g. 002.json) with the key value pairs. There are a few different ways we could test the models:

Ask each model to extract values for specific keys
Ask each model to generate a value for each of a set of keys
Ask the model to extract all keys and values from the receipt

To keep things simple, we'll go with the first option, but it would be interesting to do each and see how that affects the results.

import requests
 
indices = [str(i).zfill(3) for i in range(100)]
 
 
def load_receipt(index):
    img_path = f"https://raw.githubusercontent.com/zzzDavid/ICDAR-2019-SROIE/refs/heads/master/data/img/{index}.jpg"
    json_path = f"https://raw.githubusercontent.com/zzzDavid/ICDAR-2019-SROIE/refs/heads/master/data/key/{index}.json"
 
    json_response = requests.get(json_path).json()
    return json_response, img_path
 
 
fields, img_path = load_receipt("001")
fields

{'company': 'INDAH GIFT & HOME DECO',
 'date': '19/10/2018',
 'address': '27, JALAN DEDAP 13, TAMAN JOHOR JAYA, 81100 JOHOR BAHRU, JOHOR.',
 'total': '60.30'}

MODELS = [
    "gpt-4o",
    "gpt-4o-mini",
    "meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo",
    "meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo",
    "pixtral-12b-2409",
]
 
SYSTEM_PROMPT = """Extract the field '{key}' from the provided receipt. Return the extracted
value, and nothing else. For example, if the field is 'Total' and the value is '100',
you should just return '100'. If the field is not present, return null.
 
Do not decorate the output with any explanation, or markdown. Just return the extracted value.
"""
 
 
@braintrust.traced
async def extract_value(model, key, img_path):
    response = await client.chat.completions.create(
        model=model,
        messages=[
            {"role": "system", "content": SYSTEM_PROMPT.format(key=key)},
            {"role": "user", "content": [{"type": "image_url", "image_url": {"url": img_path}}]},
        ],
        temperature=0,
    )
    return response.choices[0].message.content.strip()
 
 
for model in MODELS:
    print("Running model: ", model)
    print(await extract_value(model, "company", img_path))
    print("\n")

Running model:  gpt-4o
INDAH GIFT & HOME DECO


Running model:  gpt-4o-mini
INDAH GIFT & HOME DECO


Running model:  meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo
60.30


Running model:  meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo
INDAH GIFT & HOME DECO


Running model:  pixtral-12b-2409
tan woon yann

Running the evaluation

Now that we were able to perform a basic sanity test, let's run an evaluation! We'll use the Levenshtein and Factuality scorers to assess performance. Levenshtein is heuristic and will tell us how closely the actual and expected strings match. Assuming some of the models will occasionally spit out superfluous explanation text, Factuality, which is LLM based, should be able to still give us an accuracy measurement.

from braintrust import Eval
 
from autoevals import Factuality, Levenshtein
 
NUM_RECEIPTS = 100
 
data = [
    {
        "input": {
            "key": key,
            "img_path": img_path,
        },
        "expected": value,
        "metadata": {
            "idx": idx,
        },
    }
    for idx, (fields, img_path) in [(idx, load_receipt(idx)) for idx in indices[:NUM_RECEIPTS]]
    for key, value in fields.items()
]
 
for model in MODELS:
 
    async def task(input):
        return await extract_value(model, input["key"], input["img_path"])
 
    await Eval(
        "Receipt Extraction",
        data=data,
        task=task,
        scores=[Levenshtein, Factuality],
        experiment_name=f"Receipt Extraction - {model}",
        metadata={"model": model},
    )

Experiment Receipt Extraction - gpt-4o is running at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20gpt-4o
Receipt Extraction [experiment_name=Receipt Extraction - gpt-4o] (data): 400it [00:00, 421962.17it/s]
Receipt Extraction [experiment_name=Receipt Extraction - gpt-4o] (tasks): 100%|██████████| 400/400 [00:42<00:00,  9.48it/s]


=========================SUMMARY=========================
84.40% 'Factuality'  score
84.93% 'Levenshtein' score

1223tok prompt_tokens
12.06tok completion_tokens
1235.06tok total_tokens

See results for Receipt Extraction - gpt-4o at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20gpt-4o

Experiment Receipt Extraction - gpt-4o-mini is running at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20gpt-4o-mini
Receipt Extraction [experiment_name=Receipt Extraction - gpt-4o-mini] (data): 400it [00:00, 76419.86it/s]
Receipt Extraction [experiment_name=Receipt Extraction - gpt-4o-mini] (tasks): 100%|██████████| 400/400 [00:41<00:00,  9.63it/s]


=========================SUMMARY=========================
Receipt Extraction - gpt-4o-mini compared to Receipt Extraction - gpt-4o:
86.81% (+01.88%) 'Levenshtein' score	(85 improvements, 48 regressions)
81.40% (-03.00%) 'Factuality'  score	(34 improvements, 42 regressions)

38052.40tok (+3682940.00%) 'prompt_tokens'    	(0 improvements, 400 regressions)
12.31tok (+25.25%) 'completion_tokens'	(62 improvements, 49 regressions)
38064.71tok (+3682965.25%) 'total_tokens'     	(0 improvements, 400 regressions)

See results for Receipt Extraction - gpt-4o-mini at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20gpt-4o-mini

Experiment Receipt Extraction - meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo is running at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20meta-llama%2FLlama-3.2-11B-Vision-Instruct-Turbo
Receipt Extraction [experiment_name=Receipt Extraction - meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo] (data): 400it [00:00, 73234.17it/s]
Receipt Extraction [experiment_name=Receipt Extraction - meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo] (tasks): 100%|██████████| 400/400 [00:26<00:00, 15.01it/s]


=========================SUMMARY=========================
Receipt Extraction - meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo compared to Receipt Extraction - gpt-4o-mini:
52.78% (-34.04%) 'Levenshtein' score	(41 improvements, 235 regressions)
56.10% (-25.30%) 'Factuality'  score	(38 improvements, 162 regressions)

89tok (-3796340.00%) 'prompt_tokens'    	(400 improvements, 0 regressions)
11.31tok (-100.50%) 'completion_tokens'	(125 improvements, 268 regressions)
100.31tok (-3796440.50%) 'total_tokens'     	(400 improvements, 0 regressions)

See results for Receipt Extraction - meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20meta-llama%2FLlama-3.2-11B-Vision-Instruct-Turbo

Experiment Receipt Extraction - meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo is running at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20meta-llama%2FLlama-3.2-90B-Vision-Instruct-Turbo
Receipt Extraction [experiment_name=Receipt Extraction - meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo] (data): 400it [00:00, 59897.24it/s]
Receipt Extraction [experiment_name=Receipt Extraction - meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo] (tasks): 100%|██████████| 400/400 [00:36<00:00, 10.90it/s]


=========================SUMMARY=========================
Receipt Extraction - meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo compared to Receipt Extraction - meta-llama/Llama-3.2-11B-Vision-Instruct-Turbo:
77.52% (+24.74%) 'Levenshtein' score	(212 improvements, 40 regressions)
79.10% (+23.00%) 'Factuality'  score	(154 improvements, 35 regressions)

89tok (-) 'prompt_tokens'    	(0 improvements, 0 regressions)
14.45tok (+313.75%) 'completion_tokens'	(75 improvements, 157 regressions)
103.45tok (+313.75%) 'total_tokens'     	(75 improvements, 157 regressions)

See results for Receipt Extraction - meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20meta-llama%2FLlama-3.2-90B-Vision-Instruct-Turbo

Experiment Receipt Extraction - pixtral-12b-2409 is running at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20pixtral-12b-2409
Receipt Extraction [experiment_name=Receipt Extraction - pixtral-12b-2409] (data): 400it [00:00, 125474.65it/s]
Receipt Extraction [experiment_name=Receipt Extraction - pixtral-12b-2409] (tasks): 100%|██████████| 400/400 [00:50<00:00,  7.88it/s]


=========================SUMMARY=========================
Receipt Extraction - pixtral-12b-2409 compared to Receipt Extraction - meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo:
66.75% (-12.35%) 'Factuality'  score	(47 improvements, 98 regressions)
73.56% (-03.96%) 'Levenshtein' score	(72 improvements, 145 regressions)

2364.51tok (+227551.00%) 'prompt_tokens'    	(0 improvements, 400 regressions)
19.22tok (+477.50%) 'completion_tokens'	(121 improvements, 252 regressions)
2383.73tok (+228028.50%) 'total_tokens'     	(0 improvements, 400 regressions)

See results for Receipt Extraction - pixtral-12b-2409 at https://www.braintrust.dev/app/braintrustdata.com/p/Receipt%20Extraction/experiments/Receipt%20Extraction%20-%20pixtral-12b-2409

Analyzing the results

Now that we have a bunch of results, let's take a look at some of the insights. If you click into the project in Braintrust, and then "Group by" model, you'll see the following:

grouped-by-model

A few quick takeaways:

it looks like gpt-4o-mini performs the best -- even slightly better than gpt-4o.
Pixtral, a 12B model, performs significantly better than LLaMa 3.2 11B and almost as well as 90B.
Both LLaMa models (for these tests, hosted on Together), are dramatically faster -- almost 3x -- than GPT-4o, GPT-4o-mini, and Pixtral.

Let's dig into these individual results in some more depth.

GPT-4o-mini vs GPT-4o

If you click into the gpt-4o experiment and compare it to gpt-4o-mini, you can drill down into the individual improvements and regressions.

Regressions

There are several different types of regressions, one of which appears to be that gpt-4o returns information in a different case than gpt-4o-mini. That may or may not be important for this use case, but if not, we could adjust our scoring functions to lowercase everything before comparing.

Casing

Pixtral vs. LLaMa 3.2

To compare Pixtral to LLaMa 3.2, you can do a multi-way comparison where the baseline is Pixtral.

Pixtral vs. LLaMa 3.2

If you filter to results where the Levenshtein score is 100%, and then drag to filter the score buckets where Levenshtein is less than 100% for LLaMa models, you'll see that 109 out of the 400 total tests match. That means that around 25% of the results had a perfect (100%) score for Pixtral and a lower score for LLaMa models.

Pixtral filter

It's useful to eyeball a few of these, where you'll see that many of the answers are just straight up incorrect for LLaMa 3.2 models.

Incorrect

Speed vs. quality trade-off

Back on the experiments page, it can be useful to view a scatterplot of score vs. duration to understand the trade-off between accuracy and speed.

Scatter plot

The LLaMa 3.2 models are significantly faster—almost 3x—without sacrificing much accuracy. For certain use cases, this can be a significant factor to consider.

Where to go from here

Now that we have some baseline evals in place, you can start to think about how to either iterate on these models to improve performance, or expand the testing to get a more comprehensive benchmark:

Try tweaking the prompt, perhaps with some few-shot examples, and see if that affects absolute and relative performance
Add a few more models into the mix and see how they perform
Dig into a few regressions and tweak the scoring methods to better reflect the actual use case

To get started with this use case in Braintrust, you can sign up for a free account and start with this Notebook. Happy evaluating!

On this page