Document & PDF Processing
Our API supports sending documents directly through the Chat Completions endpoint. The API extracts textual content from files — including OCR for PDFs with images, tables, and complex layouts — and feeds it to the model. At the moment, files must be sent in Base64 format.
Below is an example illustrating how to upload a PDF to the API:
import base64
import openai
client = openai.OpenAI(
api_key="YOUR_KEY",
base_url="https://chat.maritaca.ai/api",
)
with open("my_research_proposal.pdf", "rb") as f:
file_content = f.read()
file_content_base64 = base64.b64encode(file_content).decode("utf-8")
messages = [
{
"role": "user",
"content": [
{
"type": "file",
"file": {
"filename": "my_research.pdf",
"file_data": f"data:application/pdf;base64,{file_content_base64}",
},
},
{
"type": "text",
"text": "What is the title of the research proposed here?",
},
],
}
]
response = client.chat.completions.create(
model="sabia-4",
messages=messages,
)
Extracted content is returned in the response and can be accessed like this:
extracted_docs = response.extracted_files
for doc in extracted_docs:
print(doc["filename"]) # File name
print(doc["text"]) # Extracted text
We currently support the following file types:
- DOCX
- TXT
- CSV
- XML
- XLSX
- MD
- PNG / JPEG (images)
Images
The API also accepts images (PNG and JPEG) using the OpenAI-compatible image_url format.
Image content is extracted via OCR and passed to the model as text.
You can send an image via URL or Base64:
messages = [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": "https://example.com/photo.png",
},
},
{
"type": "text",
"text": "What does this image say?",
},
],
}
]
response = client.chat.completions.create(
model="sabia-4",
messages=messages,
)
To send the image as Base64 (no public URL needed):
with open("photo.png", "rb") as f:
img_b64 = base64.b64encode(f.read()).decode("utf-8")
messages = [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": f"data:image/png;base64,{img_b64}",
},
},
{
"type": "text",
"text": "Extract the text from this image.",
},
],
}
]
Images can also be sent as type: "file", just like PDFs and other documents:
with open("photo.png", "rb") as f:
img_b64 = base64.b64encode(f.read()).decode("utf-8")
messages = [
{
"role": "user",
"content": [
{
"type": "file",
"file": {
"filename": "photo.png",
"file_data": f"data:image/png;base64,{img_b64}",
},
},
{
"type": "text",
"text": "Extract the text from this image.",
},
],
}
]
Images are processed via OCR — the model receives extracted text, not the original image. Supported formats: PNG and JPEG.
Extraction effort
For PDFs and images, you can specify the extraction effort level. Every PDF page goes through OCR, and each image counts as one page. This setting has no effect on the other file types (DOCX, TXT, CSV, XML, XLSX and MD), which are read directly and incur no extraction charges.
If not specified, the default extraction level is medium. Values other than medium and advanced return a 400 error.
Available extraction levels
| Level | Description | Cost per page |
|---|---|---|
medium | OCR of each page: text and tables. Works best for most documents. | R$ 0,020 |
advanced | Everything in medium + text descriptions of images, charts, and flowcharts on the page. | R$ 0,025 |
You can check the extraction pricing on our platform.
Example using extraction effort
response = client.chat.completions.create(
model="sabia-4",
messages=messages,
extra_body={
"extraction_effort": "medium"
},
)
The usage field includes both token usage and the number of extracted pages:
usage = response.usage
print(usage)
# CompletionUsage(
# completion_tokens=167,
# prompt_tokens=45,
# total_tokens=212,
# extraction_details={
# "medium_effort_pages_extracted": 1,
# "advanced_effort_pages_extracted": 0,
# },
# )
Important Considerations
-
Extracted text is not stored by the API. If you plan to ask follow-up questions about the same file, store the extracted text yourself — otherwise, a new request will trigger extraction (and extraction charges for PDFs and images).
-
The combined size of all files in a request must be under 50 MB.
-
PDFs and images incur a per-page extraction charge based on the chosen level, in addition to the prompt tokens processed by the model.
-
Discounts Off-peak, Batch API, and Flex discounts apply only to input and output tokens. These discounts do not apply to the extraction cost.
Other examples
Multiple documents of different types
xml_path = "test.xml"
csv_path = "test.csv"
pdf_path = "test.pdf"
with open(xml_path, "rb") as xml_file:
xml_content = xml_file.read()
xml_content = base64.b64encode(xml_content).decode("utf-8")
with open(csv_path, "rb") as csv_file:
csv_content = csv_file.read()
csv_content = base64.b64encode(csv_content).decode("utf-8")
with open(pdf_path, "rb") as pdf_file:
pdf_content = pdf_file.read()
pdf_content = base64.b64encode(pdf_content).decode("utf-8")
csv_segment = {
"type": "file",
"file": {
"filename": "test.csv",
"file_data": f"data:text/csv;base64,{csv_content}",
},
}
xml_segment = {
"type": "file",
"file": {
"filename": "test.xml",
"file_data": f"data:application/xml;base64,{xml_content}",
},
}
pdf_segment = {
"type": "file",
"file": {
"filename": "test.pdf",
"file_data": f"data:application/pdf;base64,{pdf_content}",
},
}
messages = [
{
"role": "system",
"content": "You are a helpful assistant that can extract information from files",
},
{
"role": "user",
"content": [
csv_segment,
xml_segment,
pdf_segment,
{"type": "text", "text": "What is the content of each of these files?"},
],
},
]
r = client.chat.completions.create(
model=model_name,
messages=messages,
max_tokens=30,
extra_body={
"extraction_effort": "medium",
},
)