import warnings
from streamlit_card import card
import streamlit as st
import pandas as pd
import plotly.express as px
from xrag.config import Config
from xrag.eval.EvalModelAgent import EvalModelAgent
from xrag.eval.evaluate_rag import evaluating
from xrag.launcher import run
from xrag.eval.evaluate_rag import EvaluationResult
from xrag.process.query_transform import transform_and_query
from xrag.launcher import build_index, build_query_engine
from xrag.data.qa_loader import get_qa_dataset
AVAILABLE_METRICS = [
"NLG_chrf", "NLG_meteor", "NLG_wer", "NLG_cer", "NLG_chrf_pp",
"NLG_perplexity",
"NLG_rouge_rouge1", "NLG_rouge_rouge2", "NLG_rouge_rougeL", "NLG_rouge_rougeLsum",
"Llama_retrieval_Faithfulness", "Llama_retrieval_Relevancy", "Llama_response_correctness",
"Llama_response_semanticSimilarity", "Llama_response_answerRelevancy", "Llama_retrieval_RelevancyG",
"Llama_retrieval_FaithfulnessG",
"DeepEval_retrieval_contextualPrecision", "DeepEval_retrieval_contextualRecall",
"DeepEval_retrieval_contextualRelevancy", "DeepEval_retrieval_faithfulness",
"DeepEval_response_answerRelevancy", "DeepEval_response_hallucination",
"DeepEval_response_bias", "DeepEval_response_toxicity",
"UpTrain_Response_Completeness", "UpTrain_Response_Conciseness", "UpTrain_Response_Relevance",
"UpTrain_Response_Valid", "UpTrain_Response_Consistency", "UpTrain_Response_Response_Matching",
"UpTrain_Retrieval_Context_Relevance", "UpTrain_Retrieval_Context_Utilization",
"UpTrain_Retrieval_Factual_Accuracy", "UpTrain_Retrieval_Context_Conciseness",
"UpTrain_Retrieval_Code_Hallucination",
]
# Define options for each dropdown
LLM_OPTIONS = [
# huggingface models
# 'llama', # meta-llama/Llama-2-7b-chat-hf
# 'chatglm', # THUDM/chatglm3-6b
# 'qwen', # Qwen/Qwen1.5-7B-Chat
# 'qwen14_int8',# Qwen/Qwen1.5-14B-Chat-GPTQ-Int8
# 'qwen7_int8', # Qwen/Qwen1.5-7B-Chat-GPTQ-Int8
# 'qwen1.8', # Qwen/Qwen1.5-1.8B-Chat
# 'baichuan', # baichuan-inc/Baichuan2-7B-Chat
# 'falcon', # tiiuae/falcon-7b-instruct
# 'mpt', # mosaicml/mpt-7b-chat
# 'yi', # 01-ai/Yi-6B-Chat
'openai', # OpenAI API
'huggingface',# HuggingFace local models
'ollama', # Ollama local models
]
HF_MODEL_OPTIONS = [
'llama', # meta-llama/Llama-2-7b-chat-hf
'chatglm', # THUDM/chatglm3-6b
'qwen', # Qwen/Qwen1.5-7B-Chat
'qwen14_int8',# Qwen/Qwen1.5-14B-Chat-GPTQ-Int8
'qwen7_int8', # Qwen/Qwen1.5-7B-Chat-GPTQ-Int8
'qwen1.8', # Qwen/Qwen1.5-1.8B-Chat
'baichuan', # baichuan-inc/Baichuan2-7B-Chat
'falcon', # tiiuae/falcon-7b-instruct
'mpt', # mosaicml/mpt-7b-chat
'yi', # 01-ai/Yi-6B-Chat
]
# ollama cascade dict
OLLAMA_OPTIONS = {
"LLaMA": {
"llama2-7b": "llama2:7b",
"llama2-13b": "llama2:13b",
"llama2-70b": "llama2:70b",
"llama3.1-8b": "llama3.1:8b",
"llama3.1-70b": "llama3.1:70b",
"llama3.2-1b": "llama3.2:1b",
"llama3.2-3b": "llama3.2:3b",
},
"Mistral": {
"mistral-7b": "mistral:7b",
"mixtral": "mixtral",
"mistral-7b-q4": "mistral:7b-q4_0",
},
"others": {
"gemma-7b": "gemma:7b",
"codellama": "codellama",
"neural-chat": "neural-chat",
"other": "other"
},
}
EMBEDDING_OPTIONS = ["BAAI/bge-large-en-v1.5", "BAAI/bge-m3", "BAAI/bge-base-en-v1.5","BAAI/bge-small-en-v1.5","BAAI/bge-large-zh-v1.5","BAAI/bge-base-zh-v1.5","BAAI/bge-small-zh-v1.5"] # Add more as needed
SPLIT_TYPE_OPTIONS = ["sentence", "character", "hierarchical"]
DATASET_OPTIONS = ["hotpot_qa", "drop", "natural_questions","trivia_qa","search_qa","finqa","law"] # Replace with actual dataset options
DATASET_DISPLAY_MAP = {
"hotpot_qa": "HotpotQA",
"drop": "DropQA",
"natural_questions": "NaturalQA"
# remain for other options
}
DATASETS_INFO = {
"HotpotQA": {
"size": {
"train": "86,830",
"validation": "8,680",
"test": "968"
},
"corpus": {
"documents": "508,826",
"source": "Wikipedia"
},
"features": {
"Multi-hop": True,
"Constrained": False,
"Numerical": True,
"Set-logical": False
}
},
"DropQA": {
"size": {
"train": "78,241",
"validation": "7,824",
"test": "870"
},
"corpus": {
"documents": "6,147",
"source": "Wikipedia"
},
"features": {
"Multi-hop": True,
"Constrained": False,
"Numerical": True,
"Set-logical": True
}
},
"NaturalQA": {
"size": {
"train": "100,093",
"validation": "10,010",
"test": "1,112"
},
"corpus": {
"documents": "49,815",
"source": "Wikipedia"
},
"features": {
"Multi-hop": True,
"Constrained": True,
"Numerical": True,
"Set-logical": False
}
}
}
# frontend dataset options
FRONTEND_DATASET_OPTIONS = [DATASET_DISPLAY_MAP.get(ds, ds) for ds in DATASET_OPTIONS]
RETRIEVER_OPTIONS = ["BM25", "Vector", "Summary", "Tree", "Keyword", "Custom", "QueryFusion", "AutoMerging", "Recursive", "SentenceWindow"] # Add more as needed
POSTPROCESS_RERANK_OPTIONS = ["none","long_context_reorder", "colbertv2_rerank","bge-reranker-base"] # Add more as needed
QUERY_TRANSFORM_OPTIONS = ["none", "hyde_zeroshot", "hyde_fewshot","stepback_zeroshot","stepback_fewshot"] # Add more as needed
# follow the order listed in the docs
METRIC_DISPLAY_MAP = {
"NLG_chrf": "ChrF",
"NLG_chrf_pp": "ChrF++",
"NLG_meteor": "METEOR",
"NLG_rouge_rouge1": "ROUGE1",
"NLG_rouge_rouge2": "ROUGE2",
"NLG_rouge_rougeL": "ROUGEL",
"NLG_rouge_rougeLsum": "ROUGELSUM",
"nlg-em": "EM",
"NLG_perplexity": "PPL",
"NLG_cer": "CER",
"NLG_wer": "WER",
# Llama metrics
"Llama_retrieval_Faithfulness": "Llama-Response-Faithfulness",
"Llama_retrieval_Relevancy": "Llama-Response-Relevance",
"Llama_response_correctness": "Llama-Response-Correctness",
"Llama_response_semanticSimilarity": "Llama-Response-Similarity",
"Llama_response_answerRelevancy": "Llama-Response-Relevance++",
"Llama_retrieval_FaithfulnessG": "Llama-Response-Faithfulness+",
"Llama_retrieval_RelevancyG": "Llama-Response-Relevance+",
# UpTrain metrics
"UpTrain_Retrieval_Context_Relevance": "Uptrain-Context-Relevance",
"UpTrain_Retrieval_Context_Conciseness": "Uptrain-Context-Conciseness",
"DeepEval_response_answerRelevancy": "DeepEval-Response-Relevancy",
"UpTrain_Response_Completeness": "Uptrain-Response-Completeness",
"UpTrain_Response_Conciseness": "Uptrain-Response-Conciseness",
"UpTrain_Response_Relevance": "Uptrain-Response-Relevance",
"UpTrain_Response_Valid": "Uptrain-Response-Valid",
"UpTrain_Response_Response_Matching": "Uptrain-Response-Matching",
"UpTrain_Retrieval_Code_Hallucination": "Uptrain-Retrieval-Code-Hallucination",
# DeepEval metrics
"DeepEval_retrieval_contextualPrecision": "DeepEval-Context-Recall",
"DeepEval_retrieval_contextualRecall": "DeepEval-Context-Relevance",
"DeepEval_retrieval_contextualRelevancy": "Uptrain-Context-Consistency",
"UpTrain_Response_Consistency": "Uptrain-Context-Utilization",
"UpTrain_Retrieval_Context_Utilization": "Uptrain-Factual-Accuracy",
"UpTrain_Retrieval_Factual_Accuracy": "Uptrain-Factual-Accuracy",
"DeepEval_retrieval_faithfulness": "DeepEval-Context-Faithfulness",
"DeepEval_response_hallucination": "DeepEval-Context-Hallucination",
# others not in the docs but in the code
"DeepEval_response_bias": "DeepEval-Response-Bias",
"DeepEval_response_toxicity": "DeepEval-Response-Toxicity",
}
# frontend values
FRONTEND_AVAILABLE_METRICS = [METRIC_DISPLAY_MAP.get(metric, metric) for metric in AVAILABLE_METRICS]
# reverse map
DISPLAY_TO_BACKEND_METRIC_MAP = {v: k for k, v in METRIC_DISPLAY_MAP.items()}
[docs]
@st.cache_resource(show_spinner=False)
def get_query():
return run(cli=False)
[docs]
@st.cache_resource(show_spinner=False)
def get_qa_dataset_(dataset,files=None):
return get_qa_dataset(dataset,files=files)
[docs]
@st.cache_resource(show_spinner=False)
def get_index():
return build_index(st.session_state.qa_dataset['documents'])
[docs]
@st.cache_resource(show_spinner=False)
def get_query_engine():
return build_query_engine(st.session_state.index, st.session_state.hierarchical_storage_context)
[docs]
def main():
if "step" not in st.session_state:
st.session_state.step = 1
st.set_page_config(layout="wide")
st.title("XRAG")
cfg = Config()
if st.session_state.step == 1:
st.header("Choose your Dataset")
# find the backend dataset value
backend_dataset = cfg.dataset if cfg.dataset in DATASET_OPTIONS else DATASET_OPTIONS[0]
# find the frontend dataset value and index
frontend_dataset = DATASET_DISPLAY_MAP.get(backend_dataset, backend_dataset)
frontend_index = FRONTEND_DATASET_OPTIONS.index(frontend_dataset)
# selected_frontend_dataset = st.selectbox(
# "Dataset",
# options=FRONTEND_DATASET_OPTIONS,
# index=frontend_index,
# key="dataset"
# )
if 'dataset' not in st.session_state:
st.session_state.dataset = "HotpotQA"
cols = st.columns(3)
st.markdown("""
<style>
.stCard {
background-color: white !important;
}
.card-body {
background-color: white !important;
}
.card {
background-color: white !important;
}
/* 修改 iframe 的高度 */
iframe.stCustomComponentV1 {
height: 400px !important; /* 调整这个值来改变卡片高度 */
}
/* 减少卡片之间的间距 */
div[data-testid="column"] {
margin-top: 0 !important;
margin-bottom: 0 !important;
}
.block-container {
padding-top: 2rem !important;
padding-bottom: 2rem !important;
}
</style>
""", unsafe_allow_html=True)
for i, (dataset_name, info) in enumerate(DATASETS_INFO.items()):
with cols[i]:
is_selected = st.session_state.dataset == dataset_name
features_text = "\n".join([
f"• {k}: {'✓' if v else '✗'}"
for k, v in info['features'].items()
])
size_text = (
f"Train: {info['size']['train']}\n"
f"Val: {info['size']['validation']}\n"
f"Test: {info['size']['test']}"
)
clicked = card(
title=dataset_name,
text=(
f"📊 Dataset Size:\n{size_text}\n"
f"📚 Corpus:\n"
f"• Documents: {info['corpus']['documents']}"
# f"✨ Features:\n{features_text}"
),
image="https://hotpotqa.github.io/img/home-bg.jpg",
styles={
"card": {
"width": "100%",
"height": "300px",
"border-radius": "10px",
"box-shadow": "0 0 10px rgba(0,0,0,0.1)",
"background-color": "#ffffff", # 设置背景色为白色
"border": "4px solid #2E86C1" if is_selected else "1px solid #e0e0e0",
},
"title": {
"font-size": "20px",
"font-weight": "bold",
"color": "#ffffff", # 标题改为黑色
"margin-bottom": "10px"
},
"text": {
"font-size": "14px",
"line-height": "1.5",
"white-space": "pre-line",
"color": "#ffffff" # 文本改为黑色
}
}
)
if clicked:
st.session_state.dataset = dataset_name
st.rerun()
selected_frontend_dataset = st.session_state.dataset
# reverse map: from the frontend dataset to the backend dataset
DISPLAY_TO_BACKEND_MAP = {v: k for k, v in DATASET_DISPLAY_MAP.items()}
chosen_backend_dataset = DISPLAY_TO_BACKEND_MAP.get(selected_frontend_dataset, selected_frontend_dataset)
cfg.dataset = chosen_backend_dataset
# custom dataset
st.markdown("---")
st.markdown("## Or your own dataset")
tab1, tab2 = st.tabs(["Upload JSON", "Generate from Folder"])
with tab1:
with st.expander("About Custom Dataset"):
st.markdown("""
### Upload Custom Dataset
Please upload a JSON file containing your dataset. The JSON file should be a list of objects with the following format:
```json
[
{
"question": "What is the capital of France?",
"answer": "Paris",
"file_paths": "path/to/document.txt",
"source_text": "source of the document"
// or multiple files
// "file_paths": ["path/to/doc1.txt", "path/to/doc2.txt"]
// "source": ["source of the document 1", "source of the document 2"]
},
...
]
```
Note:
1. Make sure all file paths in the JSON are accessible from the server.
2. Supported file formats: txt, md, pdf, html, json, csv, etc.
3. Each question can reference one or multiple documents.
4. The system will automatically process and index all documents.
""")
uploaded_file = st.file_uploader("Upload your dataset", type=["json"])
with tab2:
st.markdown("""
### Generate QA Pairs from Documents
Upload a folder containing your documents, and the system will:
1. Read all documents in the folder (including subfolders)
2. Use AI to generate relevant questions and answers
3. Create a dataset in the required format
Supported file formats:
- Text files (.txt)
- Markdown files (.md)
- PDF documents (.pdf)
- HTML files (.html)
- And more...
Note:
- If the sentence length is set to -1, the system will use file level as the unit.
- If the sentence length is set to a positive number, the system will split the document into chunks of the specified length.
""")
folder_path = st.text_input("Enter folder path", value="./data/documents")
num_questions = st.number_input("Number of questions per file", min_value=1, value=3)
output_file = st.text_input("Output JSON file path", value="./data/generated_qa.json")
sentence_length = st.number_input("Sentence length", value=-1)
if st.button("Generate QA Dataset"):
try:
with st.spinner("Generating QA pairs from documents..."):
from xrag.data.qa_loader import generate_qa_from_folder
qa_pairs = generate_qa_from_folder(folder_path, output_file, num_questions, sentence_length=sentence_length)
st.success(f"Successfully generated {len(qa_pairs)} QA pairs!")
# 自动加载生成的数据集
st.session_state.qa_dataset = get_qa_dataset_("custom", output_file)
cfg.dataset = "custom"
st.session_state.step = 2
st.rerun()
except Exception as e:
st.error(f"Error generating QA dataset: {str(e)}")
if st.button("Load Dataset"):
if uploaded_file is not None:
try:
st.session_state.step = 2
with st.spinner("Loading Dataset..."):
st.session_state.qa_dataset = get_qa_dataset_("custom", uploaded_file)
cfg.dataset = "custom"
st.rerun()
except Exception as e:
st.error(f"Error loading dataset: {str(e)}")
else:
with st.spinner("Loading Dataset..."):
st.session_state.qa_dataset = get_qa_dataset_(cfg.dataset)
st.session_state.step = 2
st.rerun()
if st.session_state.step == 2:
st.header("Configure your RAG Index")
st.subheader("Settings")
cfg.llm = st.selectbox("LLM", options=LLM_OPTIONS, index=LLM_OPTIONS.index(cfg.llm) if cfg.llm in LLM_OPTIONS else 0)
if cfg.llm == "ollama":
cfg.ollama_model = st.text_input("Your Ollama Model", value=cfg.ollama_model)
elif cfg.llm == 'huggingface':
cfg.huggingface_model = st.selectbox("HuggingFace Model", options=HF_MODEL_OPTIONS, index=HF_MODEL_OPTIONS.index(cfg.huggingface_model) if cfg.huggingface_model in HF_MODEL_OPTIONS else 0)
cfg.auth_token = st.text_input("Your Auth Token", value=cfg.auth_token)
elif cfg.llm == 'openai':
cfg.api_key = st.text_input("API Key", value=cfg.api_key, type="password")
cfg.api_base = st.text_input("API Base", value=cfg.api_base)
cfg.api_name = st.text_input("Model Name", value=cfg.api_name)
st.markdown("---")
cfg.embeddings = st.selectbox("Embeddings", options=EMBEDDING_OPTIONS, index=EMBEDDING_OPTIONS.index(cfg.embeddings) if cfg.embeddings in EMBEDDING_OPTIONS else 0)
cfg.split_type = st.selectbox("Split Type", options=SPLIT_TYPE_OPTIONS, index=SPLIT_TYPE_OPTIONS.index(cfg.split_type))
cfg.chunk_size = st.number_input("Chunk Size", min_value=1, value=cfg.chunk_size, step=1)
# cfg.source_dir = st.text_input("Source Directory", value=cfg.source_dir)
cfg.persist_dir = st.text_input("Persist Directory", value=cfg.persist_dir)
# 返回或者继续
c1,c_,c2 = st.columns([1,4,1])
with c1:
if st.button("Back"):
st.session_state.step = 1
st.rerun()
with c2:
if st.button("Build Index"):
st.session_state.step = 3
with st.spinner("Building Index..."):
st.session_state.index, st.session_state.hierarchical_storage_context = get_index()
st.rerun()
# Evaluation Metrics Selection
if st.session_state.step == 3:
st.header("Configure your RAG Query Engine")
cfg.retriever = st.selectbox("Advanced Retriever", options=RETRIEVER_OPTIONS, index=RETRIEVER_OPTIONS.index(
cfg.retriever) if cfg.retriever in RETRIEVER_OPTIONS else 0)
cfg.retriever_mode = st.selectbox("Retriever Mode", options=[0, 1], index=cfg.retriever_mode)
cfg.query_transform = st.selectbox("Pre-retriever", options=QUERY_TRANSFORM_OPTIONS,
index=QUERY_TRANSFORM_OPTIONS.index(
cfg.query_transform) if cfg.query_transform in QUERY_TRANSFORM_OPTIONS else 0)
cfg.postprocess_rerank = st.selectbox("Post-process", options=POSTPROCESS_RERANK_OPTIONS,
index=POSTPROCESS_RERANK_OPTIONS.index(
cfg.postprocess_rerank) if cfg.postprocess_rerank in POSTPROCESS_RERANK_OPTIONS else 0)
cfg.text_qa_template_str = st.text_area("Text QA Template", value=cfg.text_qa_template_str)
cfg.refine_template_str = st.text_area("Refine Template", value=cfg.refine_template_str)
c1, c_, c2 = st.columns([1, 4, 1])
with c1:
if st.button("Back"):
st.session_state.step = 2
st.rerun()
with c2:
if st.button("Build Query Engine"):
st.session_state.step = 4
with st.spinner("Building Query Engine..."):
st.session_state.query_engine = get_query_engine()
st.rerun()
if st.session_state.step == 4:
st.header("Evaluate your RAG Model with single question")
prompt = st.text_input('Input your question here')
if st.button("Evaluate Your Question"):
response = transform_and_query(prompt, cfg, st.session_state.query_engine)
st.write(response.response)
# Display source text
with st.expander('Source Text'):
st.write(response.get_formatted_sources(length=1024))
c1, c_, c2 = st.columns([1, 4, 1])
with c1:
if st.button("Back"):
st.session_state.step = 3
st.rerun()
with c2:
if st.button("Evaluate Your Dataset"):
st.session_state.step = 5
st.rerun()
if st.session_state.step == 5:
st.header("Evaluate your RAG Model with your dataset")
# 确保默认选中的 metrics 按 AVAILABLE_METRICS 的顺序排列
sorted_default_metrics = [metric for metric in AVAILABLE_METRICS if metric in cfg.metrics]
sorted_frontend_default_metrics = [METRIC_DISPLAY_MAP.get(metric, metric) for metric in sorted_default_metrics]
# 用户在前端选择评测指标
selected_frontend_metrics = st.multiselect(
"Evaluation Metrics",
options=FRONTEND_AVAILABLE_METRICS,
default=sorted_frontend_default_metrics
)
# 将前端选择的显示名称映射回后端的真实名称
selected_backend_metrics = [DISPLAY_TO_BACKEND_METRIC_MAP.get(metric, metric) for metric in selected_frontend_metrics]
cfg.metrics = selected_backend_metrics
# 其他输入
cfg.test_init_total_number_documents = st.number_input(
"Total number of documents to evaluate",
min_value=1,
value=cfg.test_init_total_number_documents,
step=1
)
c1, c_, c2 = st.columns([1, 4, 1])
with c1:
if st.button("Back"):
st.session_state.step = 4
st.rerun()
with c2:
start_evaluation = st.button("Start Evaluation")
if start_evaluation:
all_num = 0
metrics = cfg.metrics.copy()
evaluateResults = EvaluationResult(metrics=metrics)
evalAgent = EvalModelAgent(cfg)
if cfg.experiment_1:
if len(st.session_state.qa_dataset) < cfg.test_init_total_number_documents:
warnings.filterwarnings('default')
warnings.warn("使用的数据集长度大于数据集本身的最大长度，请修改。 本轮代码无法运行", UserWarning)
else:
cfg.test_init_total_number_documents = cfg.n
for question, expected_answer, golden_context, golden_context_ids in zip(
st.session_state.qa_dataset['test_data']['question'][:cfg.test_init_total_number_documents],
st.session_state.qa_dataset['test_data']['expected_answer'][:cfg.test_init_total_number_documents],
st.session_state.qa_dataset['test_data']['golden_context'][:cfg.test_init_total_number_documents],
st.session_state.qa_dataset['test_data']['golden_context_ids'][:cfg.test_init_total_number_documents]
):
response = transform_and_query(question, cfg, st.session_state.query_engine)
# 返回node节点
retrieval_ids = []
retrieval_context = []
for source_node in response.source_nodes:
retrieval_ids.append(source_node.metadata['id'])
retrieval_context.append(source_node.get_content())
actual_response = response.response
eval_result = evaluating(
question, response, actual_response, retrieval_context, retrieval_ids,
expected_answer, golden_context, golden_context_ids, evaluateResults.metrics,
evalAgent
)
with st.expander(question):
st.markdown("### Answer")
st.markdown(response.response)
st.markdown('### Retrieval context')
st.markdown('\n\n'.join(retrieval_context))
st.markdown('### Expected answer')
st.markdown(expected_answer)
st.markdown('### Golden context')
st.markdown('\n\n'.join(golden_context))
print(eval_result)
evaluateResults.add(eval_result)
all_num += 1
st.markdown(evaluateResults.get_results_str())
st.success("Evaluation complete!")
st.session_state.evaluation_results = evaluateResults
# if st.session_state.step == 5:
# st.header("Evaluation Results")
# if 'evaluation_results' in st.session_state:
# results = st.session_state.evaluation_results
# display_results(results)
[docs]
def display_results(results: EvaluationResult):
# Display summary statistics
st.subheader("Summary Statistics")
summary = pd.DataFrame(results.get_summary(), index=[0])
st.table(summary)
# Display detailed metrics
st.subheader("Detailed Metrics")
metrics_df = pd.DataFrame(results.get_all_metrics())
st.dataframe(metrics_df)
# Visualize metrics
st.subheader("Metric Visualization")
metric_to_plot = st.selectbox("Select a metric to visualize", options=results.metrics)
fig = px.box(metrics_df, y=metric_to_plot, title=f"Distribution of {metric_to_plot}")
st.plotly_chart(fig)
# Display sample evaluations
st.subheader("Sample Evaluations")
num_samples = st.slider("Number of samples to display", min_value=1, max_value=len(results.evaluations), value=5)
for i, eval_result in enumerate(results.evaluations[:num_samples]):
st.write(f"Sample {i+1}")
st.json(eval_result)
if __name__ == "__main__":
main()