def format_element(
    element: Element, keep_refs: bool = False, latex_env: bool = False
) -> List[str]:
 """
    Formats a given Element into a list of formatted strings.

    Args:
        element (Element): The element to be formatted.
        keep_refs (bool, optional): Whether to keep references in the formatting. Default is False.
        latex_env (bool, optional): Whether to use LaTeX environment formatting. Default is False.

    Returns:
        List[str]: A list of formatted strings representing the formatted element.
    """
 ...
 ...
 if isinstance(element, Table):
        parts = [
 "[TABLE%s]\n\begin{table}\n"
 % (str(uuid4())[:5] if element.id is None else ":" + str(element.id))
 ]
        parts.extend(format_children(element, keep_refs, latex_env))
        caption_parts = format_element(element.caption, keep_refs, latex_env)
        remove_trailing_whitespace(caption_parts)
        parts.append("\end{table}\n")
 if len(caption_parts) > 0:
            parts.extend(caption_parts + ["\n"])
        parts.append("[ENDTABLE]\n\n")
 return parts
 ...
 ...

pip install langchain
pip install chromadb
pip install nougat-ocr

langchain                                0.1.12
langchain-community                      0.0.28
langchain-core                           0.1.31
langchain-openai                         0.0.8
langchain-text-splitters                 0.0.1

chroma-hnswlib                           0.7.3
chromadb                                 0.4.24

nougat-ocr                               0.1.17

import os
os.environ["OPENAI_API_KEY"] = "YOUR_OPEN_AI_KEY"

import subprocess
import uuid

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langchain.retrievers.multi_vector import MultiVectorRetriever
from langchain.storage import InMemoryStore
from langchain_community.vectorstores import Chroma
from langchain_core.documents import Document
from langchain_openai import OpenAIEmbeddings
from langchain_core.runnables import RunnablePassthrough

def june_run_nougat(file_path, output_dir):
 # Run Nougat and store results as Mathpix Markdown
    cmd = ["nougat", file_path, "-o", output_dir, "-m", "0.1.0-base", "--no-skipping"]
    res = subprocess.run(cmd) 
 if res.returncode != 0:
 print("Error when running nougat.")
 return res.returncode
 else:
 print("Operation Completed!")
 return 0

def june_get_tables_from_mmd(mmd_path):
    f = open(mmd_path)
    lines = f.readlines()
    res = []
    tmp = []
    flag = ""
 for line in lines:
 if line == "\begin{table}\n":
            flag = "BEGINTABLE"
 elif line == "\end{table}\n":
            flag = "ENDTABLE"
 
 if flag == "BEGINTABLE":
            tmp.append(line)
 elif flag == "ENDTABLE":
            tmp.append(line)
            flag = "CAPTION"
 elif flag == "CAPTION":
            tmp.append(line)
            flag = "MARKDOWN"
 print('-' * 100)
 print(''.join(tmp))
            res.append(''.join(tmp))
            tmp = []

 return res

file_path = "YOUR_PDF_PATH"
output_dir = "YOUR_OUTPUT_DIR_PATH"

if june_run_nougat(file_path, output_dir) == 1:
 import sys
    sys.exit(1)

mmd_path = output_dir + '/' + os.path.splitext(file_path)[0].split('/')[-1] + ".mmd" 
tables = june_get_tables_from_mmd(mmd_path)

Operation Completed!
----------------------------------------------------------------------------------------------------
\begin{table}
\begin{tabular}{l c c c} \hline \hline Layer Type & Complexity per Layer & Sequential Operations & Maximum Path Length \ \hline Self-Attention & (O(n^{2}\cdot d)) & (O(1)) & (O(1)) \ Recurrent & (O(n\cdot d^{2})) & (O(n)) & (O(n)) \ Convolutional & (O(k\cdot n\cdot d^{2})) & (O(1)) & (O(log_{k}(n))) \ Self-Attention (restricted) & (O(r\cdot n\cdot d)) & (O(1)) & (O(n/r)) \ \hline \hline \end{tabular}
\end{table}
Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. (n) is the sequence length, (d) is the representation dimension, (k) is the kernel size of convolutions and (r) the size of the neighborhood in restricted self-attention.

----------------------------------------------------------------------------------------------------
\begin{table}
\begin{tabular}{l c c c c} \hline \hline \multirow{2}{*}{Model} & \multicolumn{2}{c}{BLEU} & \multicolumn{2}{c}{Training Cost (FLOPs)} \ \cline{2-5}  & EN-DE & EN-FR & EN-DE & EN-FR \ \hline ByteNet [18] & 23.75 & & & \ Deep-Att + PosUnk [39] & & 39.2 & & (1.0\cdot 10^{20}) \ GNMT + RL [38] & 24.6 & 39.92 & (2.3\cdot 10^{19}) & (1.4\cdot 10^{20}) \ ConvS2S [9] & 25.16 & 40.46 & (9.6\cdot 10^{18}) & (1.5\cdot 10^{20}) \ MoE [32] & 26.03 & 40.56 & (2.0\cdot 10^{19}) & (1.2\cdot 10^{20}) \ \hline Deep-Att + PosUnk Ensemble [39] & & 40.4 & & (8.0\cdot 10^{20}) \ GNMT + RL Ensemble [38] & 26.30 & 41.16 & (1.8\cdot 10^{20}) & (1.1\cdot 10^{21}) \ ConvS2S Ensemble [9] & 26.36 & **41.29** & (7.7\cdot 10^{19}) & (1.2\cdot 10^{21}) \ \hline Transformer (base model) & 27.3 & 38.1 & & (\mathbf{3.3\cdot 10^{18}}) \ Transformer (big) & **28.4** & **41.8** & & (2.3\cdot 10^{19}) \ \hline \hline \end{tabular}
\end{table}
Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the English-to-German and English-to-French newstest2014 tests at a fraction of the training cost.

----------------------------------------------------------------------------------------------------
\begin{table}
\begin{tabular}{c|c c c c c c c c|c c c c} \hline \hline  & (N) & (d_{\text{model}}) & (d_{\text{ff}}) & (h) & (d_{k}) & (d_{v}) & (P_{drop}) & (\epsilon_{ls}) & train steps & PPL & BLEU & params \ \hline base & 6 & 512 & 2048 & 8 & 64 & 64 & 0.1 & 0.1 & 100K & 4.92 & 25.8 & 65 \ \hline \multirow{4}{*}{(A)} & \multicolumn{1}{c}{} & & 1 & 512 & 512 & & & & 5.29 & 24.9 & \  & & & & 4 & 128 & 128 & & & & 5.00 & 25.5 & \  & & & & 16 & 32 & 32 & & & & 4.91 & 25.8 & \  & & & & 32 & 16 & 16 & & & & 5.01 & 25.4 & \ \hline (B) & \multicolumn{1}{c}{} & & \multicolumn{1}{c}{} & & 16 & & & & & 5.16 & 25.1 & 58 \  & & & & & 32 & & & & & 5.01 & 25.4 & 60 \ \hline \multirow{4}{*}{(C)} & 2 & \multicolumn{1}{c}{} & & & & & & & & 6.11 & 23.7 & 36 \  & 4 & & & & & & & & 5.19 & 25.3 & 50 \  & 8 & & & & & & & & 4.88 & 25.5 & 80 \  & & 256 & & 32 & 32 & & & & 5.75 & 24.5 & 28 \  & 1024 & & 128 & 128 & & & & 4.66 & 26.0 & 168 \  & & 1024 & & & & & & 5.12 & 25.4 & 53 \  & & 4096 & & & & & & 4.75 & 26.2 & 90 \ \hline \multirow{4}{*}{(D)} & \multicolumn{1}{c}{} & & & & & 0.0 & & 5.77 & 24.6 & \  & & & & & & 0.2 & & 4.95 & 25.5 & \  & & & & & & & 0.0 & 4.67 & 25.3 & \  & & & & & & & 0.2 & 5.47 & 25.7 & \ \hline (E) & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & & \multicolumn{1}{c}{} & & & & & 4.92 & 25.7 & \ \hline big & 6 & 1024 & 4096 & 16 & & 0.3 & 300K & **4.33** & **26.4** & 213 \ \hline \hline \end{tabular}
\end{table}
Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base model. All metrics are on the English-to-German translation development set, newstest2013. Listed perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to per-word perplexities.

----------------------------------------------------------------------------------------------------
\begin{table}
\begin{tabular}{c|c|c} \hline
**Parser** & **Training** & **WSJ 23 F1** \ \hline Vinyals & Kaiser et al. (2014) [37] & WSJ only, discriminative & 88.3 \ Petrov et al. (2006) [29] & WSJ only, discriminative & 90.4 \ Zhu et al. (2013) [40] & WSJ only, discriminative & 90.4 \ Dyer et al. (2016) [8] & WSJ only, discriminative & 91.7 \ \hline Transformer (4 layers) & WSJ only, discriminative & 91.3 \ \hline Zhu et al. (2013) [40] & semi-supervised & 91.3 \ Huang & Harper (2009) [14] & semi-supervised & 91.3 \ McClosky et al. (2006) [26] & semi-supervised & 92.1 \ Vinyals & Kaiser el al. (2014) [37] & semi-supervised & 92.1 \ \hline Transformer (4 layers) & semi-supervised & 92.7 \ \hline Luong et al. (2015) [23] & multi-task & 93.0 \ Dyer et al. (2016) [8] & generative & 93.3 \ \hline \end{tabular}
\end{table}
Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)* [5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. _CoRR_, abs/1406.1078, 2014.

# Prompt
prompt_text = """You are an assistant tasked with summarizing tables and text. \ 
Give a concise summary of the table or text. The table is formatted in LaTeX, and its caption is in plain text format: {element}  """
prompt = ChatPromptTemplate.from_template(prompt_text)

# Summary chain
model = ChatOpenAI(temperature = 0, model = "gpt-3.5-turbo")
summarize_chain = {"element": lambda x: x} | prompt | model | StrOutputParser()
# Get table summaries
table_summaries = summarize_chain.batch(tables, {"max_concurrency": 5})
print(table_summaries)

# The vectorstore to use to index the child chunks
vectorstore = Chroma(collection_name = "summaries", embedding_function = OpenAIEmbeddings())

# The storage layer for the parent documents
store = InMemoryStore()
id_key = "doc_id"

# The retriever (empty to start)
retriever = MultiVectorRetriever(
    vectorstore = vectorstore,
    docstore = store,
    id_key = id_key,
    search_kwargs={"k": 1} # Solving Number of requested results 4 is greater than number of elements in index..., updating n_results = 1
)

# Add tables
table_ids = [str(uuid.uuid4()) for _ in tables]
summary_tables = [
    Document(page_content = s, metadata = {id_key: table_ids[i]})
 for i, s in enumerate(table_summaries)
]
retriever.vectorstore.add_documents(summary_tables)
retriever.docstore.mset(list(zip(table_ids, tables)))

# Prompt template
template = """Answer the question based only on the following context, which can include text and tables, there is a table in LaTeX format and a table caption in plain text format:
{context}
Question: {question}
"""
prompt = ChatPromptTemplate.from_template(template)

# LLM
model = ChatOpenAI(temperature = 0, model = "gpt-3.5-turbo")


# Simple RAG pipeline
chain = (
 {"context": retriever, "question": RunnablePassthrough()}
 | prompt
 | model
 | StrOutputParser()
)


print(chain.invoke("when layer type is Self-Attention, what is the Complexity per Layer?")) # Query about table 1

print(chain.invoke("Which parser performs worst for BLEU EN-DE")) # Query about table 2

print(chain.invoke("Which parser performs best for WSJ 23 F1")) # Query about table 4

import os
os.environ["OPENAI_API_KEY"] = "YOUR_OPEN_AI_KEY"

import subprocess
import uuid

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langchain.retrievers.multi_vector import MultiVectorRetriever
from langchain.storage import InMemoryStore
from langchain_community.vectorstores import Chroma
from langchain_core.documents import Document
from langchain_openai import OpenAIEmbeddings
from langchain_core.runnables import RunnablePassthrough


def june_run_nougat(file_path, output_dir):
 # Run Nougat and store results as Mathpix Markdown
    cmd = ["nougat", file_path, "-o", output_dir, "-m", "0.1.0-base", "--no-skipping"]
    res = subprocess.run(cmd) 
 if res.returncode != 0:
 print("Error when running nougat.")
 return res.returncode
 else:
 print("Operation Completed!")
 return 0

def june_get_tables_from_mmd(mmd_path):
    f = open(mmd_path)
    lines = f.readlines()
    res = []
    tmp = []
    flag = ""
 for line in lines:
 if line == "\begin{table}\n":
            flag = "BEGINTABLE"
 elif line == "\end{table}\n":
            flag = "ENDTABLE"
 
 if flag == "BEGINTABLE":
            tmp.append(line)
 elif flag == "ENDTABLE":
            tmp.append(line)
            flag = "CAPTION"
 elif flag == "CAPTION":
            tmp.append(line)
            flag = "MARKDOWN"
 print('-' * 100)
 print(''.join(tmp))
            res.append(''.join(tmp))
            tmp = []

 return res

file_path = "YOUR_PDF_PATH"
output_dir = "YOUR_OUTPUT_DIR_PATH"

if june_run_nougat(file_path, output_dir) == 1:
 import sys
    sys.exit(1)

mmd_path = output_dir + '/' + os.path.splitext(file_path)[0].split('/')[-1] + ".mmd" 
tables = june_get_tables_from_mmd(mmd_path)


# Prompt
prompt_text = """You are an assistant tasked with summarizing tables and text. \ 
Give a concise summary of the table or text. The table is formatted in LaTeX, and its caption is in plain text format: {element}  """
prompt = ChatPromptTemplate.from_template(prompt_text)

# Summary chain
model = ChatOpenAI(temperature = 0, model = "gpt-3.5-turbo")
summarize_chain = {"element": lambda x: x} | prompt | model | StrOutputParser()
# Get table summaries
table_summaries = summarize_chain.batch(tables, {"max_concurrency": 5})
print(table_summaries)

# The vectorstore to use to index the child chunks
vectorstore = Chroma(collection_name = "summaries", embedding_function = OpenAIEmbeddings())

# The storage layer for the parent documents
store = InMemoryStore()
id_key = "doc_id"

# The retriever (empty to start)
retriever = MultiVectorRetriever(
    vectorstore = vectorstore,
    docstore = store,
    id_key = id_key,
    search_kwargs={"k": 1} # Solving Number of requested results 4 is greater than number of elements in index..., updating n_results = 1
)

# Add tables
table_ids = [str(uuid.uuid4()) for _ in tables]
summary_tables = [
    Document(page_content = s, metadata = {id_key: table_ids[i]})
 for i, s in enumerate(table_summaries)
]
retriever.vectorstore.add_documents(summary_tables)
retriever.docstore.mset(list(zip(table_ids, tables)))


# Prompt template
template = """Answer the question based only on the following context, which can include text and tables, there is a table in LaTeX format and a table caption in plain text format:
{context}
Question: {question}
"""
prompt = ChatPromptTemplate.from_template(template)

# LLM
model = ChatOpenAI(temperature = 0, model = "gpt-3.5-turbo")

# Simple RAG pipeline
chain = (
 {"context": retriever, "question": RunnablePassthrough()}
 | prompt
 | model
 | StrOutputParser()
)

print(chain.invoke("when layer type is Self-Attention, what is the Complexity per Layer?")) # Query about table 1

print(chain.invoke("Which parser performs worst for BLEU EN-DE")) # Query about table 2

print(chain.invoke("Which parser performs best for WSJ 23 F1")) # Query about table 4