Advanced Document Parsing For Enterprises

1
%%capture
2
! sudo apt install tesseract-ocr poppler-utils
3
! pip install "cohere<5" fsspec hnswlib google-cloud-documentai google-cloud-storage boto3 langchain-text-splitters llama_parse pytesseract pdf2image pandas

1
data_dir = "data/document-parsing"
2
source_filename = "fda-approved-drug"
3
extension = "pdf"

1
from pathlib import Path
2
3
sources = ["gcp", "aws", "unstructured-io", "llamaparse-text", "llamaparse-markdown", "pytesseract"]
4
5
filenames = ["{}-parsed-fda-approved-drug.txt".format(source) for source in sources]
6
filenames.append("fda-approved-drug.pdf")
7
8
for filename in filenames:
9
    file_path = Path(f"{data_dir}/{filename}")
10
    if file_path.is_file() == False:
11
        print(f"File {filename} not found at {data_dir}!")

1
def store_document(path: str, doc_content: str):
2
    with open(path, 'w') as f:
3
      f.write(doc_content)

1
import json
2
3
def insert_citations_in_order(text, citations, documents):
4
    """
5
    A helper function to pretty print citations.
6
    """
7
8
    citations_reference = {}
9
    for index, doc in enumerate(documents):
10
        citations_reference[index] = doc
11
12
    offset = 0
13
    # Process citations in the order they were provided
14
    for citation in citations:
15
        # Adjust start/end with offset
16
        start, end = citation['start'] + offset, citation['end'] + offset
17
        citation_numbers = []
18
        for doc_id in citation["document_ids"]:
19
            for citation_index, doc in citations_reference.items():
20
                if doc["id"] == doc_id:
21
                    citation_numbers.append(citation_index)
22
        references = "(" + ", ".join("[{}]".format(num) for num in citation_numbers) + ")"
23
        modification = f'{text[start:end]} {references}'
24
        # Replace the cited text with its bolded version + placeholder
25
        text = text[:start] + modification + text[end:]
26
        # Update the offset for subsequent replacements
27
        offset += len(modification) - (end - start)
28
29
    # Add the citations at the bottom of the text
30
    text_with_citations = f'{text}'
31
    citations_reference = ["[{}]: {}".format(x["id"], x["text"]) for x in citations_reference.values()]
32
33
    return text_with_citations, "\n".join(citations_reference)

1
def format_docs_for_chat(documents):
2
  return [{"id": str(index), "text": x} for index, x in enumerate(documents)]

1
"""
2
Extracted from https://cloud.google.com/document-ai/docs/samples/documentai-batch-process-document
3
"""
4
5
import re
6
from typing import Optional
7
8
from google.api_core.client_options import ClientOptions
9
from google.api_core.exceptions import InternalServerError
10
from google.api_core.exceptions import RetryError
11
from google.cloud import documentai  # type: ignore
12
from google.cloud import storage
13
14
project_id = ""
15
location = ""
16
processor_id = ""
17
gcs_output_uri = ""
18
# credentials_file = "populate if you are running in a non Vertex AI environment."
19
gcs_input_prefix = ""
20
21
22
def batch_process_documents(
23
    project_id: str,
24
    location: str,
25
    processor_id: str,
26
    gcs_output_uri: str,
27
    gcs_input_prefix: str,
28
    timeout: int = 400
29
) -> None:
30
    parsed_documents = []
31
32
    # Client configs
33
    opts = ClientOptions(api_endpoint=f"{location}-documentai.googleapis.com")
34
    # With credentials
35
    # opts = ClientOptions(api_endpoint=f"{location}-documentai.googleapis.com", credentials_file=credentials_file)
36
37
    client = documentai.DocumentProcessorServiceClient(client_options=opts)
38
    processor_name = client.processor_path(project_id, location, processor_id)
39
40
    # Input storage configs
41
    gcs_prefix = documentai.GcsPrefix(gcs_uri_prefix=gcs_input_prefix)
42
    input_config = documentai.BatchDocumentsInputConfig(gcs_prefix=gcs_prefix)
43
44
    # Output storage configs
45
    gcs_output_config = documentai.DocumentOutputConfig.GcsOutputConfig(gcs_uri=gcs_output_uri, field_mask=None)
46
    output_config = documentai.DocumentOutputConfig(gcs_output_config=gcs_output_config)
47
    storage_client = storage.Client()
48
    # With credentials
49
    # storage_client = storage.Client.from_service_account_json(json_credentials_path=credentials_file)
50
51
    # Batch process docs request
52
    request = documentai.BatchProcessRequest(
53
        name=processor_name,
54
        input_documents=input_config,
55
        document_output_config=output_config,
56
    )
57
58
    # batch_process_documents returns a long running operation
59
    operation = client.batch_process_documents(request)
60
61
    # Continually polls the operation until it is complete.
62
    # This could take some time for larger files
63
    try:
64
        print(f"Waiting for operation {operation.operation.name} to complete...")
65
        operation.result(timeout=timeout)
66
    except (RetryError, InternalServerError) as e:
67
        print(e.message)
68
69
    # Get output document information from completed operation metadata
70
    metadata = documentai.BatchProcessMetadata(operation.metadata)
71
    if metadata.state != documentai.BatchProcessMetadata.State.SUCCEEDED:
72
        raise ValueError(f"Batch Process Failed: {metadata.state_message}")
73
74
    print("Output files:")
75
    # One process per Input Document
76
    for process in list(metadata.individual_process_statuses):
77
        matches = re.match(r"gs://(.*?)/(.*)", process.output_gcs_destination)
78
        if not matches:
79
            print("Could not parse output GCS destination:", process.output_gcs_destination)
80
            continue
81
82
        output_bucket, output_prefix = matches.groups()
83
        output_blobs = storage_client.list_blobs(output_bucket, prefix=output_prefix)
84
85
        # Document AI may output multiple JSON files per source file
86
        # (Large documents get split in multiple file "versions" doc --> parsed_doc_0 + parsed_doc_1 ...)
87
        for blob in output_blobs:
88
            # Document AI should only output JSON files to GCS
89
            if blob.content_type != "application/json":
90
                print(f"Skipping non-supported file: {blob.name} - Mimetype: {blob.content_type}")
91
                continue
92
93
            # Download JSON file as bytes object and convert to Document Object
94
            print(f"Fetching {blob.name}")
95
            document = documentai.Document.from_json(blob.download_as_bytes(), ignore_unknown_fields=True)
96
            # Store the filename and the parsed versioned document content as a tuple
97
            parsed_documents.append((blob.name.split("/")[-1].split(".")[0], document.text))
98
99
    print("Finished document parsing process.")
100
    return parsed_documents
101
102
# Call service
103
# versioned_parsed_documents = batch_process_documents(
104
#     project_id=project_id,
105
#     location=location,
106
#     processor_id=processor_id,
107
#     gcs_output_uri=gcs_output_uri,
108
#     gcs_input_prefix=gcs_input_prefix
109
# )

1
"""
2
Post process parsed document and store it locally.
3
Make sure to run this in a Google Vertex AI environment or include a credentials file.
4
"""
5
6
"""
7
from pathlib import Path
8
from collections import defaultdict
9
10
parsed_documents = []
11
combined_versioned_parsed_documents = defaultdict(list)
12
13
# Assemble versioned documents together ({"doc_name": [(0, doc_content_0), (1, doc_content_1), ...]}).
14
for filename, doc_content in versioned_parsed_documents:
15
  filename, version = "-".join(filename.split("-")[:-1]), filename.split("-")[-1]
16
  combined_versioned_parsed_documents[filename].append((version, doc_content))
17
18
# Sort documents by version and join the content together.
19
for filename, docs in combined_versioned_parsed_documents.items():
20
  doc_content = " ".join([x[1] for x in sorted(docs, key=lambda x: x[0])])
21
  parsed_documents.append((filename, doc_content))
22
23
# Store parsed documents in local storage.
24
for filename, doc_content in parsed_documents:
25
 file_path = "{}/{}-parsed-{}.txt".format(data_dir, "gcp", source_filename)
26
 store_document(file_path, doc_content)
27
"""

1
filename = "gcp-parsed-{}.txt".format(source_filename)
2
with open("{}/{}".format(data_dir, filename), "r") as doc:
3
    parsed_document = doc.read()
4
5
print(parsed_document[:1000])

1
# source: https://github.com/awsdocs/aws-doc-sdk-examples/tree/main/python/example_code/textract
2
3
# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
4
# SPDX-License-Identifier: Apache-2.0
5
6
"""
7
Purpose
8
9
Shows how to use the AWS SDK for Python (Boto3) with Amazon Textract to
10
detect text, form, and table elements in document images.
11
"""
12
13
import json
14
import logging
15
from botocore.exceptions import ClientError
16
17
logger = logging.getLogger(__name__)
18
19
20
# snippet-start:[python.example_code.textract.TextractWrapper]
21
class TextractWrapper:
22
    """Encapsulates Textract functions."""
23
24
    def __init__(self, textract_client, s3_resource, sqs_resource):
25
        """
26
        :param textract_client: A Boto3 Textract client.
27
        :param s3_resource: A Boto3 Amazon S3 resource.
28
        :param sqs_resource: A Boto3 Amazon SQS resource.
29
        """
30
        self.textract_client = textract_client
31
        self.s3_resource = s3_resource
32
        self.sqs_resource = sqs_resource
33
34
    # snippet-end:[python.example_code.textract.TextractWrapper]
35
36
    # snippet-start:[python.example_code.textract.DetectDocumentText]
37
    def detect_file_text(self, *, document_file_name=None, document_bytes=None):
38
        """
39
        Detects text elements in a local image file or from in-memory byte data.
40
        The image must be in PNG or JPG format.
41
42
        :param document_file_name: The name of a document image file.
43
        :param document_bytes: In-memory byte data of a document image.
44
        :return: The response from Amazon Textract, including a list of blocks
45
                 that describe elements detected in the image.
46
        """
47
        if document_file_name is not None:
48
            with open(document_file_name, "rb") as document_file:
49
                document_bytes = document_file.read()
50
        try:
51
            response = self.textract_client.detect_document_text(
52
                Document={"Bytes": document_bytes}
53
            )
54
            logger.info("Detected %s blocks.", len(response["Blocks"]))
55
        except ClientError:
56
            logger.exception("Couldn't detect text.")
57
            raise
58
        else:
59
            return response
60
61
    # snippet-end:[python.example_code.textract.DetectDocumentText]
62
63
    # snippet-start:[python.example_code.textract.AnalyzeDocument]
64
    def analyze_file(
65
        self, feature_types, *, document_file_name=None, document_bytes=None
66
    ):
67
        """
68
        Detects text and additional elements, such as forms or tables, in a local image
69
        file or from in-memory byte data.
70
        The image must be in PNG or JPG format.
71
72
        :param feature_types: The types of additional document features to detect.
73
        :param document_file_name: The name of a document image file.
74
        :param document_bytes: In-memory byte data of a document image.
75
        :return: The response from Amazon Textract, including a list of blocks
76
                 that describe elements detected in the image.
77
        """
78
        if document_file_name is not None:
79
            with open(document_file_name, "rb") as document_file:
80
                document_bytes = document_file.read()
81
        try:
82
            response = self.textract_client.analyze_document(
83
                Document={"Bytes": document_bytes}, FeatureTypes=feature_types
84
            )
85
            logger.info("Detected %s blocks.", len(response["Blocks"]))
86
        except ClientError:
87
            logger.exception("Couldn't detect text.")
88
            raise
89
        else:
90
            return response
91
92
    # snippet-end:[python.example_code.textract.AnalyzeDocument]
93
94
    # snippet-start:[python.example_code.textract.helper.prepare_job]
95
    def prepare_job(self, bucket_name, document_name, document_bytes):
96
        """
97
        Prepares a document image for an asynchronous detection job by uploading
98
        the image bytes to an Amazon S3 bucket. Amazon Textract must have permission
99
        to read from the bucket to process the image.
100
101
        :param bucket_name: The name of the Amazon S3 bucket.
102
        :param document_name: The name of the image stored in Amazon S3.
103
        :param document_bytes: The image as byte data.
104
        """
105
        try:
106
            bucket = self.s3_resource.Bucket(bucket_name)
107
            bucket.upload_fileobj(document_bytes, document_name)
108
            logger.info("Uploaded %s to %s.", document_name, bucket_name)
109
        except ClientError:
110
            logger.exception("Couldn't upload %s to %s.", document_name, bucket_name)
111
            raise
112
113
    # snippet-end:[python.example_code.textract.helper.prepare_job]
114
115
    # snippet-start:[python.example_code.textract.helper.check_job_queue]
116
    def check_job_queue(self, queue_url, job_id):
117
        """
118
        Polls an Amazon SQS queue for messages that indicate a specified Textract
119
        job has completed.
120
121
        :param queue_url: The URL of the Amazon SQS queue to poll.
122
        :param job_id: The ID of the Textract job.
123
        :return: The status of the job.
124
        """
125
        status = None
126
        try:
127
            queue = self.sqs_resource.Queue(queue_url)
128
            messages = queue.receive_messages()
129
            if messages:
130
                msg_body = json.loads(messages[0].body)
131
                msg = json.loads(msg_body["Message"])
132
                if msg.get("JobId") == job_id:
133
                    messages[0].delete()
134
                    status = msg.get("Status")
135
                    logger.info(
136
                        "Got message %s with status %s.", messages[0].message_id, status
137
                    )
138
            else:
139
                logger.info("No messages in queue %s.", queue_url)
140
        except ClientError:
141
            logger.exception("Couldn't get messages from queue %s.", queue_url)
142
        else:
143
            return status
144
145
    # snippet-end:[python.example_code.textract.helper.check_job_queue]
146
147
    # snippet-start:[python.example_code.textract.StartDocumentTextDetection]
148
    def start_detection_job(
149
        self, bucket_name, document_file_name, sns_topic_arn, sns_role_arn
150
    ):
151
        """
152
        Starts an asynchronous job to detect text elements in an image stored in an
153
        Amazon S3 bucket. Textract publishes a notification to the specified Amazon SNS
154
        topic when the job completes.
155
        The image must be in PNG, JPG, or PDF format.
156
157
        :param bucket_name: The name of the Amazon S3 bucket that contains the image.
158
        :param document_file_name: The name of the document image stored in Amazon S3.
159
        :param sns_topic_arn: The Amazon Resource Name (ARN) of an Amazon SNS topic
160
                              where the job completion notification is published.
161
        :param sns_role_arn: The ARN of an AWS Identity and Access Management (IAM)
162
                             role that can be assumed by Textract and grants permission
163
                             to publish to the Amazon SNS topic.
164
        :return: The ID of the job.
165
        """
166
        try:
167
            response = self.textract_client.start_document_text_detection(
168
                DocumentLocation={
169
                    "S3Object": {"Bucket": bucket_name, "Name": document_file_name}
170
                },
171
                NotificationChannel={
172
                    "SNSTopicArn": sns_topic_arn,
173
                    "RoleArn": sns_role_arn,
174
                },
175
            )
176
            job_id = response["JobId"]
177
            logger.info(
178
                "Started text detection job %s on %s.", job_id, document_file_name
179
            )
180
        except ClientError:
181
            logger.exception("Couldn't detect text in %s.", document_file_name)
182
            raise
183
        else:
184
            return job_id
185
186
    # snippet-end:[python.example_code.textract.StartDocumentTextDetection]
187
188
    # snippet-start:[python.example_code.textract.GetDocumentTextDetection]
189
    def get_detection_job(self, job_id):
190
        """
191
        Gets data for a previously started text detection job.
192
193
        :param job_id: The ID of the job to retrieve.
194
        :return: The job data, including a list of blocks that describe elements
195
                 detected in the image.
196
        """
197
        try:
198
            response = self.textract_client.get_document_text_detection(JobId=job_id)
199
            job_status = response["JobStatus"]
200
            logger.info("Job %s status is %s.", job_id, job_status)
201
        except ClientError:
202
            logger.exception("Couldn't get data for job %s.", job_id)
203
            raise
204
        else:
205
            return response
206
207
    # snippet-end:[python.example_code.textract.GetDocumentTextDetection]
208
209
    # snippet-start:[python.example_code.textract.StartDocumentAnalysis]
210
    def start_analysis_job(
211
        self,
212
        bucket_name,
213
        document_file_name,
214
        feature_types,
215
        sns_topic_arn,
216
        sns_role_arn,
217
    ):
218
        """
219
        Starts an asynchronous job to detect text and additional elements, such as
220
        forms or tables, in an image stored in an Amazon S3 bucket. Textract publishes
221
        a notification to the specified Amazon SNS topic when the job completes.
222
        The image must be in PNG, JPG, or PDF format.
223
224
        :param bucket_name: The name of the Amazon S3 bucket that contains the image.
225
        :param document_file_name: The name of the document image stored in Amazon S3.
226
        :param feature_types: The types of additional document features to detect.
227
        :param sns_topic_arn: The Amazon Resource Name (ARN) of an Amazon SNS topic
228
                              where job completion notification is published.
229
        :param sns_role_arn: The ARN of an AWS Identity and Access Management (IAM)
230
                             role that can be assumed by Textract and grants permission
231
                             to publish to the Amazon SNS topic.
232
        :return: The ID of the job.
233
        """
234
        try:
235
            response = self.textract_client.start_document_analysis(
236
                DocumentLocation={
237
                    "S3Object": {"Bucket": bucket_name, "Name": document_file_name}
238
                },
239
                NotificationChannel={
240
                    "SNSTopicArn": sns_topic_arn,
241
                    "RoleArn": sns_role_arn,
242
                },
243
                FeatureTypes=feature_types,
244
            )
245
            job_id = response["JobId"]
246
            logger.info(
247
                "Started text analysis job %s on %s.", job_id, document_file_name
248
            )
249
        except ClientError:
250
            logger.exception("Couldn't analyze text in %s.", document_file_name)
251
            raise
252
        else:
253
            return job_id
254
255
    # snippet-end:[python.example_code.textract.StartDocumentAnalysis]
256
257
    # snippet-start:[python.example_code.textract.GetDocumentAnalysis]
258
    def get_analysis_job(self, job_id):
259
        """
260
        Gets data for a previously started detection job that includes additional
261
        elements.
262
263
        :param job_id: The ID of the job to retrieve.
264
        :return: The job data, including a list of blocks that describe elements
265
                 detected in the image.
266
        """
267
        try:
268
            response = self.textract_client.get_document_analysis(JobId=job_id)
269
            job_status = response["JobStatus"]
270
            logger.info("Job %s status is %s.", job_id, job_status)
271
        except ClientError:
272
            logger.exception("Couldn't get data for job %s.", job_id)
273
            raise
274
        else:
275
            return response
276
277
278
# snippet-end:[python.example_code.textract.GetDocumentAnalysis]

1
import boto3
2
3
textract_client = boto3.client('textract')
4
s3_client = boto3.client('s3')
5
6
textractWrapper = TextractWrapper(textract_client, s3_client, None)

1
bucket_name = "your-bucket-name"
2
sns_topic_arn = "your-sns-arn" # this can be found under the topic you created in the Amazon SNS dashboard
3
sns_role_arn = "sns-role-arn" # this is an IAM role that allows Textract to interact with SNS
4
5
file_name = "fda-approved-drug.pdf"

1
# kick off a text detection job. This returns a job ID.
2
job_id = textractWrapper.start_detection_job(bucket_name=bucket_name, document_file_name=file_name,
3
                                    sns_topic_arn=sns_topic_arn, sns_role_arn=sns_role_arn)

1
def get_text_results_from_textract(job_id):
2
    response = textract_client.get_document_text_detection(JobId=job_id)
3
    collection_of_textract_responses = []
4
    pages = [response]
5
6
    collection_of_textract_responses.append(response)
7
8
    while 'NextToken' in response:
9
        next_token = response['NextToken']
10
        response = textract_client.get_document_text_detection(JobId=job_id, NextToken=next_token)
11
        pages.append(response)
12
        collection_of_textract_responses.append(response)
13
    return collection_of_textract_responses
14
15
def get_the_text_with_required_info(collection_of_textract_responses):
16
    total_text = []
17
    total_text_with_info = []
18
    running_sequence_number = 0
19
20
    font_sizes_and_line_numbers = {}
21
    for page in collection_of_textract_responses:
22
        per_page_text = []
23
        blocks = page['Blocks']
24
        for block in blocks:
25
            if block['BlockType'] == 'LINE':
26
                block_text_dict = {}
27
                running_sequence_number += 1
28
                block_text_dict.update(text=block['Text'])
29
                block_text_dict.update(page=block['Page'])
30
                block_text_dict.update(left_indent=round(block['Geometry']['BoundingBox']['Left'], 2))
31
                font_height = round(block['Geometry']['BoundingBox']['Height'], 3)
32
                line_number = running_sequence_number
33
                block_text_dict.update(font_height=round(block['Geometry']['BoundingBox']['Height'], 3))
34
                block_text_dict.update(indent_from_top=round(block['Geometry']['BoundingBox']['Top'], 2))
35
                block_text_dict.update(text_width=round(block['Geometry']['BoundingBox']['Width'], 2))
36
                block_text_dict.update(line_number=running_sequence_number)
37
38
                if font_height in font_sizes_and_line_numbers:
39
                    line_numbers = font_sizes_and_line_numbers[font_height]
40
                    line_numbers.append(line_number)
41
                    font_sizes_and_line_numbers[font_height] = line_numbers
42
                else:
43
                    line_numbers = []
44
                    line_numbers.append(line_number)
45
                    font_sizes_and_line_numbers[font_height] = line_numbers
46
47
                total_text.append(block['Text'])
48
                per_page_text.append(block['Text'])
49
                total_text_with_info.append(block_text_dict)
50
51
    return total_text, total_text_with_info, font_sizes_and_line_numbers
52
53
def get_text_with_line_spacing_info(total_text_with_info):
54
    i = 1
55
    text_info_with_line_spacing_info = []
56
    while (i < len(total_text_with_info) - 1):
57
        previous_line_info = total_text_with_info[i - 1]
58
        current_line_info = total_text_with_info[i]
59
        next_line_info = total_text_with_info[i + 1]
60
        if current_line_info['page'] == next_line_info['page'] and previous_line_info['page'] == current_line_info[
61
            'page']:
62
            line_spacing_after = round((next_line_info['indent_from_top'] - current_line_info['indent_from_top']), 2)
63
            spacing_with_prev = round((current_line_info['indent_from_top'] - previous_line_info['indent_from_top']), 2)
64
            current_line_info.update(line_space_before=spacing_with_prev)
65
            current_line_info.update(line_space_after=line_spacing_after)
66
            text_info_with_line_spacing_info.append(current_line_info)
67
        else:
68
            text_info_with_line_spacing_info.append(None)
69
        i += 1
70
    return text_info_with_line_spacing_info

1
all_text = "\n".join([line["text"] if line else "" for line in text_info_with_line_spacing])
2
3
with open(f"aws-parsed-{source_filename}.txt", "w") as f:
4
  f.write(all_text)

1
filename = "aws-parsed-{}.txt".format(source_filename)
2
with open("{}/{}".format(data_dir, filename), "r") as doc:
3
    parsed_document = doc.read()
4
5
print(parsed_document[:1000])

1
import os
2
import requests
3
4
UNSTRUCTURED_URL = "" # enter service endpoint, for example "http://localhost:9500/general/v0/general" (assuming the container is running locally and exposing the service with a -p 9500:9500 port mapping)
5
6
7
parsed_documents = []
8
9
input_path = "{}/{}.{}".format(data_dir, source_filename, extension)
10
with open(input_path, 'rb') as file_data:
11
    response = requests.post(
12
        url=UNSTRUCTURED_URL,
13
        files={"files": ("{}.{}".format(source_filename, extension), file_data)},
14
        data={
15
            "output_format": (None, "application/json"),
16
            "strategy": "fast",
17
            "pdf_infer_table_structure": "true",
18
            "include_page_breaks": "true"
19
        },
20
        headers={"Accept": "application/json"}
21
    )
22
23
parsed_response = response.json()
24
25
parsed_document = " ".join([parsed_entry["text"] for parsed_entry in parsed_response])
26
print("Parsed {}".format(source_filename))

1
"""
2
Post process parsed document and store it locally.
3
"""
4
5
file_path = "{}/{}-parsed-fda-approved-drug.txt".format(data_dir, "unstructured-io")
6
store_document(file_path, parsed_document)

1
filename = "unstructured-io-parsed-{}.txt".format(source_filename)
2
with open("{}/{}".format(data_dir, filename), "r") as doc:
3
    parsed_document = doc.read()
4
5
print(parsed_document[:1000])

1
import os
2
from llama_parse import LlamaParse
3
4
import nest_asyncio # needed to notebook env
5
nest_asyncio.apply() # needed to notebook env
6
7
llama_index_api_key = "{API_KEY}"
8
input_path = "{}/{}.{}".format(data_dir, source_filename, extension)

1
# Text mode
2
text_parser = LlamaParse(
3
    api_key=llama_index_api_key,
4
    result_type="text"
5
)
6
7
text_response = text_parser.load_data(input_path)
8
text_parsed_document = " ".join([parsed_entry.text for parsed_entry in text_response])
9
10
print("Parsed {} to text".format(source_filename))

1
"""
2
Post process parsed document and store it locally.
3
"""
4
5
file_path = "{}/{}-text-parsed-fda-approved-drug.txt".format(data_dir, "llamaparse")
6
store_document(file_path, text_parsed_document)

1
# Markdown mode
2
markdown_parser = LlamaParse(
3
    api_key=llama_index_api_key,
4
    result_type="markdown"
5
)
6
7
markdown_response = markdown_parser.load_data(input_path)
8
markdown_parsed_document = " ".join([parsed_entry.text for parsed_entry in markdown_response])
9
10
print("Parsed {} to markdown".format(source_filename))

1
"""
2
Post process parsed document and store it locally.
3
"""
4
5
file_path = "{}/{}-markdown-parsed-fda-approved-drug.txt".format(data_dir, "llamaparse")
6
store_document(file_path, markdown_parsed_document)

1
# Text parsing
2
3
filename = "llamaparse-text-parsed-{}.txt".format(source_filename)
4
5
with open("{}/{}".format(data_dir, filename), "r") as doc:
6
    parsed_document = doc.read()
7
8
print(parsed_document[:1000])

1
# Markdown parsing
2
3
filename = "llamaparse-markdown-parsed-fda-approved-drug.txt"
4
with open("{}/{}".format(data_dir, filename), "r") as doc:
5
    parsed_document = doc.read()
6
7
print(parsed_document[:1000])

1
from matplotlib import pyplot as plt
2
from pdf2image import convert_from_path
3
import pytesseract

1
# pdf2image extracts as a list of PIL.Image objects
2
pages = convert_from_path(filename)

1
# we look at the first page as a sanity check:
2
3
plt.imshow(pages[0])
4
plt.axis('off')
5
plt.show()

1
label_ocr_pytesseract = "".join([pytesseract.image_to_string(page) for page in pages])

1
print(label_ocr_pytesseract[:200])

HIGHLIGHTS OF PRESCRIBING INFORMATION
These highlights do not include all the information needed to use
IWILFIN™ safely and effectively. See full prescribing information for
IWILFIN.
IWILFIN™ (eflor

1
label_ocr_pytesseract = "".join([pytesseract.image_to_string(page) for page in pages])
2
3
with open(f"pytesseract-parsed-{source_filename}.txt", "w") as f:
4
  f.write(label_ocr_pytesseract)

1
filename = "pytesseract-parsed-{}.txt".format(source_filename)
2
with open("{}/{}".format(data_dir, filename), "r") as doc:
3
    parsed_document = doc.read()
4
5
print(parsed_document[:1000])

1
import cohere
2
co = cohere.Client(api_key="{API_KEY}")

1
"""
2
Document Questions
3
"""
4
prompt = "What are the most common adverse reactions of Iwilfin?"
5
# prompt = "What is the recommended dosage of Iwilfin on body surface area between 0.5 m2 and 0.75 m2?"
6
# prompt = "I need a succinct summary of the compound name, indication, route of administration, and mechanism of action of Iwilfin."
7
8
"""
9
Choose one of the above solutions
10
"""
11
source = "gcp"
12
# source = "aws"
13
# source = "unstructured-io"
14
# source = "llamaparse-text"
15
# source = "llamaparse-markdown"
16
# source = "pytesseract"

1
"""
2
Read parsed document content and chunk data
3
"""
4
5
import os
6
from langchain_text_splitters import RecursiveCharacterTextSplitter
7
8
documents = []
9
10
with open("{}/{}-parsed-fda-approved-drug.txt".format(data_dir, source), "r") as doc:
11
    doc_content = doc.read()
12
13
"""
14
Personal notes on chunking
15
https://medium.com/@ayhamboucher/llm-based-context-splitter-for-large-documents-445d3f02b01b
16
"""
17
18
19
# Chunk doc content
20
text_splitter = RecursiveCharacterTextSplitter(
21
    chunk_size=512,
22
    chunk_overlap=200,
23
    length_function=len,
24
    is_separator_regex=False
25
)
26
27
# Split the text into chunks with some overlap
28
chunks_ = text_splitter.create_documents([doc_content])
29
documents = [c.page_content for c in chunks_]
30
31
print("Source document has been broken down to {} chunks".format(len(documents)))

1
"""
2
Embed document chunks
3
"""
4
document_embeddings = co.embed(texts=documents, model="embed-v4.0", input_type="search_document").embeddings

1
"""
2
Create document index and add embedded chunks
3
"""
4
5
import hnswlib
6
7
index = hnswlib.Index(space='ip', dim=1536) # space: inner product
8
index.init_index(max_elements=len(document_embeddings), ef_construction=512, M=64)
9
index.add_items(document_embeddings, list(range(len(document_embeddings))))
10
print("Count:", index.element_count)

Count: 115

1
"""
2
Embed search query
3
Fetch k nearest neighbors
4
"""
5
6
query_emb = co.embed(texts=[prompt], model='embed-v4.0', input_type="search_query").embeddings
7
default_knn = 10
8
knn = default_knn if default_knn <= index.element_count else index.element_count
9
result = index.knn_query(query_emb, k=knn)
10
neighbors = [(result[0][0][i], result[1][0][i]) for i in range(len(result[0][0]))]
11
relevant_docs = [documents[x[0]] for x in sorted(neighbors, key=lambda x: x[1])]

1
"""
2
Rerank retrieved documents
3
"""
4
5
rerank_results = co.rerank(query=prompt, documents=relevant_docs, top_n=3, model='rerank-english-v2.0').results
6
reranked_relevant_docs = format_docs_for_chat([x.document["text"] for x in rerank_results])

1
"""
2
Call the /chat endpoint with command-a
3
"""
4
5
response = co.chat(
6
    message=prompt,
7
    model="command-a-03-2025",
8
    documents=reranked_relevant_docs
9
)
10
11
cited_response, citations_reference = insert_citations_in_order(response.text, response.citations, reranked_relevant_docs)
12
print(cited_response)
13
print("\n")
14
print("References:")
15
print(citations_reference)

1
import pandas as pd
2
results = pd.read_csv("{}/results-table.csv".format(data_dir))

1
question = input("""
2
Question 1: What are the most common adverse reactions of Iwilfin?
3
Question 2: What is the recommended dosage of Iwilfin on body surface area between 0.5 m2 and 0.75 m2?
4
Question 3: I need a succinct summary of the compound name, indication, route of administration, and mechanism of action of Iwilfin.
5
6
Pick which question you want to see (1,2,3):  """)
7
references = input("Do you want to see the references as well? References are long and noisy (y/n): ")
8
print("\n\n")
9
10
index = {"1": 0, "2": 3, "3": 6}[question]
11
12
for src in ["gcp", "aws", "unstructured-io", "llamaparse-text", "llamaparse-markdown", "pytesseract"]:
13
  print("| {} |".format(src))
14
  print("\n")
15
  print(results[src][index])
16
  if references == "y":
17
    print("\n")
18
    print("References:")
19
    print(results[src][index+1])
20
  print("\n")

Question 1: What are the most common adverse reactions of Iwilfin?
Question 2: What is the recommended dosage of Iwilfin on body surface area between 0.5 m2 and 0.75 m2?
Question 3: I need a succinct summary of the compound name, indication, route of administration, and mechanism of action of Iwilfin.
Pick which question you want to see (1,2,3):  3
Do you want to see the references as well? References are long and noisy (y/n): n
| gcp |
Compound Name: eflornithine hydrochloride ([0], [1], [2]) (IWILFIN ([1])™)
Indication: used to reduce the risk of relapse in adult and paediatric patients with high-risk neuroblastoma (HRNB) ([1], [3]), who have responded at least partially to prior multiagent, multimodality therapy. ([1], [3], [4])
Route of Administration: IWILFIN™ tablets ([1], [3], [4]) are taken orally twice daily ([3], [4]), with doses ranging from 192 to 768 mg based on body surface area. ([3], [4])
Mechanism of Action: IWILFIN™ is an ornithine decarboxylase inhibitor. ([0], [2])
| aws |
Compound Name: eflornithine ([0], [1], [2], [3]) (IWILFIN ([0])™)
Indication: used to reduce the risk of relapse ([0], [3]) in adults ([0], [3]) and paediatric patients ([0], [3]) with high-risk neuroblastoma (HRNB) ([0], [3]) who have responded to prior therapies. ([0], [3], [4])
Route of Administration: Oral ([2], [4])
Mechanism of Action: IWILFIN is an ornithine decarboxylase inhibitor. ([1])
| unstructured-io |
Compound Name: Iwilfin ([1], [2], [3], [4]) (eflornithine) ([0], [2], [3], [4])
Indication: Iwilfin is indicated to reduce the risk of relapse ([1], [3]) in adult and paediatric patients ([1], [3]) with high-risk neuroblastoma (HRNB) ([1], [3]), who have responded to prior anti-GD2 ([1]) immunotherapy ([1], [4]) and multi-modality therapy. ([1])
Route of Administration: Oral ([0], [3])
Mechanism of Action: Iwilfin is an ornithine decarboxylase inhibitor. ([1], [2], [3], [4])
| llamaparse-text |
Compound Name: IWILFIN ([2], [3]) (eflornithine) ([3])
Indication: IWILFIN is used to reduce the risk of relapse ([1], [2], [3]) in adult and paediatric patients ([1], [2], [3]) with high-risk neuroblastoma (HRNB) ([1], [2], [3]), who have responded at least partially to certain prior therapies. ([2], [3])
Route of Administration: IWILFIN is administered as a tablet. ([2])
Mechanism of Action: IWILFIN is an ornithine decarboxylase inhibitor. ([0], [1], [4])
| llamaparse-markdown |
Compound Name: IWILFIN ([1], [2]) (eflornithine) ([1])
Indication: IWILFIN is indicated to reduce the risk of relapse ([1], [2]) in adult and paediatric patients ([1], [2]) with high-risk neuroblastoma (HRNB) ([1], [2]), who have responded at least partially ([1], [2], [3]) to prior anti-GD2 immunotherapy ([1], [2]) and multiagent, multimodality therapy. ([1], [2], [3])
Route of Administration: Oral ([0], [1], [3], [4])
Mechanism of Action: IWILFIN acts as an ornithine decarboxylase inhibitor. ([1])
| pytesseract |
Compound Name: IWILFIN™ ([0], [2]) (eflornithine) ([0], [2])
Indication: IWILFIN is indicated to reduce the risk of relapse ([0], [2]) in adult and paediatric patients ([0], [2]) with high-risk neuroblastoma (HRNB) ([0], [2]), who have responded positively to prior anti-GD2 immunotherapy and multiagent, multimodality therapy. ([0], [2], [4])
Route of Administration: IWILFIN is administered orally ([0], [1], [3], [4]), in the form of a tablet. ([1])
Mechanism of Action: IWILFIN acts as an ornithine decarboxylase inhibitor. ([0])